Skip to content

Mini Shai-Hulud npm/PyPI worm campaign

Summary

Mini Shai-Hulud is the 2026 continuation of the Shai-Hulud npm worm tradecraft: malicious package installs execute inside developer or CI/CD environments, steal credentials, and use reachable publishing or repository access to spread into additional packages and repositories.

Public reporting from Wiz, Snyk, Akamai, JFrog, Socket, Unit 42, and Microsoft describes multiple waves in April-May 2026. Several vendors attribute the later waves to TeamPCP or describe them as TeamPCP-linked; keep that attribution caveated unless a firsthand operator statement or official source is being cited directly.

Page role

This is the canonical operation page for the broad Mini Shai-Hulud npm/PyPI campaign. Keep actor-level judgments on TeamPCP and narrow April 22 Bitwarden/Checkmarx distribution-lane details on Bitwarden / Checkmarx Shai-Hulud Third Coming campaign.

Tags

Why this matters

  • The May 2026 TanStack wave showed that a malicious npm release can carry valid provenance/attestation when the legitimate release workflow is hijacked before publication.
  • The campaign targets the places with the most blast radius: developer laptops, GitHub Actions runners, npm trusted-publishing workflows, cloud credentials, Kubernetes secrets, and package maintainer accounts.
  • Removing a compromised package is not enough if the payload already created persistence, poisoned repository files, or extracted tokens that can publish more packages.

Publicly reported wave sequence

September-November 2025: original Shai-Hulud waves

  • Microsoft and Unit 42 describe the original Shai-Hulud activity as a self-propagating npm worm era shift.
  • The campaign pattern: run during package install, steal tokens/secrets, create public exfiltration repositories, and use npm access to republish infected packages.
  • Later 2025 waves added more automation and destructive behavior.
  • Socket's July 16, 2026 writeup, citing 404 Media reporting, shows the long tail of those original waves: a threat actor using the handle ellie.191 reportedly said they reached Suno after one employee was compromised by Shai-Hulud, then used harvested GitHub and cloud credentials to access source code, a customer list, and Stripe-related payment data. Suno told 404 Media the incident occurred in November 2025, was limited and quickly contained, primarily involved outdated source code, and did not expose full credit card numbers stored in Stripe. Socket's caveat is important for attribution: the actor may have planted the trojanized packages, or may simply have found Shai-Hulud-exfiltrated credentials in the victim-created public GitHub repository and reused them. For defenders, the durable point is that Shai-Hulud public credential dumps can become downstream corporate breaches months later, even when the original package incident is no longer active.

March 2026: TeamPCP / Trivy precursor activity

  • Wiz and other researchers reported a TeamPCP-linked Trivy compromise that used GitHub Actions/release infrastructure to steal runner secrets and publish malicious artifacts.
  • This established several recurring motifs later seen in Mini Shai-Hulud coverage: runner memory scraping, cloud/Kubernetes credential harvesting, package/release trust abuse, encrypted exfiltration, and fallback GitHub repository exfiltration.
  • Wiz CIRT's post-compromise reporting shows why stolen-token handling needs to extend beyond package cleanup: observed TeamPCP activity validated cloud and SaaS secrets with TruffleHog-style API calls within hours, enumerated AWS IAM/EC2/Lambda/RDS/Route 53/S3/ECS/Secrets Manager, abused GitHub PATs to create malicious workflow pull requests and delete workflow logs, cloned repositories at scale, used ECS Exec / SSM paths for code execution inside containers, and pulled data from S3, databases, Secrets Manager, and source repositories.
  • FortiGuard Labs' June 26 incident report provides a concrete CI identity to production-data case study. In an AWS estate affected by aggressive Shai-Hulud activity in early May, investigators later observed the Jenkins EC2 instance role used from external addresses, creation of an administrator user named cloudops-monitor, security-group and database-control-plane changes, mass Secrets Manager enumeration, repeated Redshift Data API queries, S3 exfiltration-policy staging, SSM SendCommand, and SES identity-verification activity. FortiGuard says leadership confirmed a warehouse breach, but explicitly states that host forensics have not definitively proved which malicious package led to the cloud compromise; temporal alignment, the stolen Jenkins-role credential class, and known worm capabilities strongly support the suspected link. Treat this as high-confidence evidence of the consequence chain, not definitive package-to-incident attribution.
  • The earliest cloud signal was use of EC2 instance-profile credentials from outside AWS. FortiGuard correlated IMDS access on the Jenkins runner and outbound traffic to 89.22.231[.]63:8080 with later privileged activity from the same address; 185.204.1[.]225 performed the first external Jenkins-role reconnaissance and escalation. Durable cloud pivots include IAM user cloudops-monitor, inline policies exfil-s3-write and exfil-s3-full, STS session names exfil, exfil10, and exfil12, unusual ModifyDBCluster / ModifyCluster calls, Redshift ingress on TCP/5439, high-volume ListSecrets / GetSecretValue, GetClusterCredentials, and roughly 90 ExecuteStatement / result cycles.
  • See also: Trivy → TeamPCP → CanisterWorm timeline.

April 22, 2026: Bitwarden / Checkmarx "Third Coming" wave

  • Unit 42 reported a TeamPCP-attributed wave that included malicious @bitwarden/cli@2026.4.0 and public GitHub artifacts containing the string Shai-Hulud: The Third Coming.
  • The same payload family reportedly appeared across Checkmarx distribution lanes: poisoned checkmarx/kics Docker Hub images, checkmarx/ast-github-action, and Checkmarx VS Code extensions.
  • Unit 42's reported @bitwarden/cli package used both a preinstall hook and a bw command-name masquerade path via bw_setup.js, giving the malware a secondary execution path even when install scripts were blocked.
  • This wave adds a durable defender lesson for later Mini Shai-Hulud triage: package-registry compromise, IDE-extension compromise, Docker image poisoning, and GitHub Actions abuse can all be lanes for the same payload and credential-theft infrastructure.
  • See also: Bitwarden / Checkmarx Shai-Hulud Third Coming campaign.

April 29-30, 2026: SAP / Intercom / PyPI / Packagist expansion

  • Wiz reported Mini Shai-Hulud-style malicious versions in SAP ecosystem npm packages including @cap-js/sqlite, @cap-js/postgres, @cap-js/db-service, and mbt.
  • The same reporting later added intercom-client and PyPI lightning packages as related compromises under analysis.
  • Reported behavior included preinstall execution, Bun-based loaders, obfuscated JavaScript payloads, cloud/GitHub/npm/Kubernetes/Vault credential harvesting, Russian locale guardrails, and GitHub-based encrypted exfiltration.
  • Wiz assessed TeamPCP responsibility with high confidence based on shared cryptographic material and implementation overlaps, while noting that references to older Shai-Hulud operations do not by themselves prove a single operator across every wave.
  • Socket's SAP CAP / Cloud MTA analysis adds operationally useful detail: mbt@1.2.48, @cap-js/db-service@2.10.1, @cap-js/postgres@2.2.2, and @cap-js/sqlite@2.2.2 were published in a short April 29 window with injected package.json, setup.mjs, and large obfuscated execution.js files. The loader downloaded Bun 1.3.13 from GitHub Releases when Bun was absent, ran the payload under Bun, and deleted the temporary runtime directory. Socket estimated meaningful SAP ecosystem reach at roughly 52k weekly downloads for mbt, 10k for @cap-js/postgres, 260k for @cap-js/db-service, and 250k for @cap-js/sqlite.
  • Socket reported the SAP payload family mirrored Bitwarden/Checkmarx cryptographic tooling, including a PBKDF2 / XOR string-protection layer, while generating unique deployment keys per payload. On developer hosts it harvested SSH keys, cloud credentials, Kubernetes config, Docker/Git/npm material, environment files, AI-tool configs such as Claude/Kiro MCP settings, crypto wallets, messaging app data, shell history, GitHub CLI tokens, and cloud metadata credentials; on CI runners it attempted Runner.Worker memory scraping to recover masked GitHub Actions secrets.
  • Socket reported that intercom-client@7.0.4 introduced setup.mjs plus an 11.7 MB router_runtime.js payload, ran during npm preinstall, downloaded Bun from GitHub without integrity checks, harvested Kubernetes/Vault/cloud/developer secrets, and exfiltrated through GitHub infrastructure.
  • Socket's follow-up expanded the Intercom chain into Packagist: intercom/intercom-php@5.0.2 was replaced by force-updated tag metadata and converted into a Composer plugin via composer-plugin-api, src/composerPlugin.php, post-install-cmd / post-update-cmd, and setup-intercom.sh, which downloaded Bun 1.3.13 and executed the same router_runtime.js-style payload.
  • Intercom told Socket the root cause was a local install of pyannote-audio that pulled the compromised PyPI lightning dependency, linking a PyPI foothold to the npm intercom-client compromise and then to the Packagist intercom/intercom-php artifact. This is a durable ecosystem-expansion lesson: Mini Shai-Hulud-style activity can move from a developer endpoint into multiple package registries through local dependency installs, repository access, and mutable package metadata rather than through one registry's native publishing flow alone.

May 11-12, 2026: TanStack and trusted-publishing abuse

  • Snyk reported malicious artifacts across @tanstack/* packages published by the legitimate TanStack release pipeline after attacker-controlled code hijacked the runner mid-workflow.
  • Unit 42 later quantified the initial TanStack burst as 84 malicious artifacts across 42 @tanstack/* packages within six minutes, expanding by end of day to 373 malicious versions across 169 npm packages plus compromised PyPI packages. Unit 42 estimated roughly 520 million cumulative downloads during the affected window.
  • Snyk’s key point: SLSA provenance can prove where the artifact was built, but it does not prove the runtime workflow was clean if attacker-controlled code executed before publication. Unit 42 called this the first documented case of a worm publishing malicious npm packages with valid SLSA Build Level 3 provenance.
  • Akamai, JFrog, and Unit 42 describe the chain as privileged workflow abuse: a pull_request_target workflow checked out fork-controlled code, a poisoned pnpm cache was written with a precomputed release cache key, a legitimate release workflow later restored that state, and the payload extracted GitHub Actions OIDC material from Runner.Worker memory to obtain npm publishing credentials.
  • Unit 42 reported that the malicious TanStack packages used an injected optionalDependencies reference to an orphaned commit surfaced under the legitimate fork network, while secondary propagation victims such as UiPath, Mistral AI, and OpenSearch reverted to more familiar preinstall execution.
  • Unit 42 also warned that the May 11 payload installed a background service that polled api.github.com/user with the stolen token and, if the token was revoked while the daemon was active, executed destructive home-directory deletion. This makes containment order especially important: stop active execution and isolate hosts before broad token revocation when this variant may be running.
  • Akamai and Unit 42 reported that weaponized Mini Shai-Hulud source code appeared publicly on GitHub after the TanStack wave, increasing copycat risk and weakening attribution based only on worm lineage. Socket separately reported that TeamPCP and BreachForums promoted a Shai-Hulud supply-chain attack contest, creating an explicit incentive for lower-tier copycats to target package ecosystems by download count.
  • OX Security reported a concrete copycat cluster after the source-code leak: four typosquatted npm packages (chalk-tempalte, @deadcode09284814/axios-util, axois-utils, and color-style-utils) from one publisher, with chalk-tempalte described as an almost unmodified Shai-Hulud clone. OX reported infostealer behavior across the set, one DDoS botnet-style phantom bot package, C2 / exfiltration endpoints including 87e0bbc636999b[.]lhr[.]life, 80[.]200[.]28[.]28:2222, b94b6bcfa27554[.]lhr[.]life, and edcf8b03c84634[.]lhr[.]life, and GitHub dead-drop repository strings such as A Mini Sha1-Hulud has Appeared. This reinforces that copycat triage should not assume TeamPCP authorship merely because the payload lineage or strings match the leaked worm.
  • On May 27, CISA added CVE-2026-45321 to KEV for the TanStack incident, describing malicious @tanstack/* releases published under a trusted identity and setting a June 10 remediation due date for covered agencies. Treat this as public exploitation validation for the trusted-publishing / cache-poisoning lane, not as a normal application bug.

April-May 2026: broader npm/PyPI spread

  • On September 7, Aikido reported the 111-day resurfacing of the May 19 AntV-wave payload: four packages (feishu-docx-mcp@0.3.2, bmc-i18n-extract-cli@1.1.1, blueai-cli@0.7.0, bmc-translate-utils@1.1.1) published within the same hour by the same npm account carried a root-level index.js byte-identical to the May 19 @antv payload (SHA256 e37e3ddeeaaa9e0c4fdbcb829b4895a6521031c80053fc436625b61e6ee5b1a6, preinstall → bun run index.js, .vscode/tasks.json / .claude/settings.json persistence, C2 t[.]m-kosche[.]com, Dune-themed dead-drop repositories with reversed "Shai-Hulud" description strings). Aikido's detection history shows 319 package versions with that hash, all first seen May 19, zero hits May 20 through September 6, then these four on September 7, a 111-day dormancy-to-reactivation gap Aikido calls the longest it has observed from a Shai-Hulud payload. The durable lesson is about registry-level scanning, not just the worm: npm has run publish-time malware scanning (a 5-15 minute pre-publish hold) since July 2026, and a hash-identical reactivation of a publicly fingerprinted, internationally covered payload is a lookup the scanner should have caught as the floor of its claimed coverage. Keep attribution caveated: no self-identifying campaign marker was recovered, and public tooling plus copycat reuse remain plausible. Triage: hunt the SHA256, the four package names, and Dune-themed repositories; do not treat this as evidence of new operator tooling.
  • On August 4, StepSecurity, Socket, and Aikido reported ChainDrop, an active npm worm that began with the high-impact keyv / cacheable maintainer ecosystem and rapidly spread across unrelated scopes. The wave adds setup.mjs and Math_Symbol.js plus a preinstall hook, downloads Bun 1.3.13, steals npm/GitHub/cloud/Vault/Kubernetes and runner credentials, and republishes reachable packages with stolen npm tokens or OIDC trusted-publishing credentials. Socket also found GitHub and DNS exfiltration plus .claude/settings.json SessionStart and .vscode/tasks.json folderOpen hooks that execute from a cloned source tree without an npm install. Aikido reported at least 1,280 affected packages by its 13:20 CEST update; treat all package counts and lists as fast-moving. The tradecraft strongly matches Shai-Hulud, but Socket did not recover a self-identifying campaign marker and public tooling makes copycat reuse possible, so do not infer TeamPCP attribution from lineage alone. See ChainDrop keyv / cacheable npm worm for live-response pivots, hashes, and the initial package set.
  • Unit 42's July 15 npm threat-landscape update adds the July 14 AsyncAPI miasma-train-p1 wave to the Mini Shai-Hulud / Miasma lineage. Unit 42 reported five trojanized AsyncAPI packages, release-pipeline compromise across four core AsyncAPI GitHub repositories, and a process gap where protected main branches coexisted with unprotected pre-production release branches such as next and schema that could still publish. The follow-up adds two durable defender lessons to this canonical campaign page: runtime import by AI/editor tooling can launch a poisoned package without a manual npm install, and Miasma-family operators are adding rate-limited propagation controls (maxGen = 4, canary rollout starting at 5%) rather than behaving like uncontrolled worms. Microsoft's same-day root-cause analysis adds a concrete pull_request_target pwn-request path that exposed the asyncapi-bot PAT, let unauthorized commits reach auto-publish branches, and produced valid OIDC provenance under npm-oidc-no-reply@github[.]com; it also reports Defender detections Trojan:JS/MiasmStealer.SC and Trojan:Script/Supychain.A. See AsyncAPI generator / specs Miasma compromise for indicators and response guidance.
  • On June 1, Wiz and StepSecurity reported Miasma, a new Mini Shai-Hulud-derived wave affecting the @redhat-cloud-services npm namespace. Wiz counted at least 32 compromised package releases with roughly 80,000 combined weekly downloads and said most malicious versions had been revoked by its 13:00 UTC update, with two still remaining at publication time; OX later counted 31+ affected packages, 116,282 accumulated weekly downloads, and more than 210 GitHub repositories with stolen credentials. StepSecurity's analysis of @redhat-cloud-services/host-inventory-client@5.0.3 found a 4.2 MB obfuscated index.js executed by a preinstall hook, published through GitHub Actions OIDC from RedHatInsights/javascript-clients, indicating upstream CI/CD pipeline compromise rather than a one-off tarball edit.
  • Wiz's later root-cause update reported that a compromised GitHub account pushed malicious orphan commits into RedHatInsights/frontend-components, RedHatInsights/javascript-clients, and RedHatInsights/platform-frontend-ai-toolkit in two waves. Those commits added a minimal release workflow triggered on any branch with id-token: write, pinned actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd, ran an obfuscated Bun payload, requested npm OIDC identity, and published package versions with valid provenance / SLSA attestations. This keeps the TanStack lesson alive: valid provenance can prove a compromised workflow produced the artifact, not that the artifact is safe.
  • Boost Security's analysis of the Red Hat wave highlights a trust-model gap rather than a payload change: npm trusted publishing checked the GitHub organization, repository, workflow filename, and optional environment, but not the branch/ref. Because no protected deployment environment constrained the publisher, a counterfeit ci.yml, ci.yaml, or release.yml on a throwaway branch could mint an accepted OIDC token, while Sigstore provenance truthfully recorded the untrusted branch without blocking publication. Boost notes PyPI's trusted-publishing model has the same optional-environment pattern, making deployment-environment branch rules a cross-registry hardening control rather than an npm-only lesson.
  • The Miasma payload keeps the Mini Shai-Hulud install-time credential-harvesting pattern but changes branding and collection scope: Wiz says it is derived from publicly released Mini Shai-Hulud malware, replaces Dune references with Greek-mythology / spartan theming, creates repositories with the description Miasma: The Spreading Blight, and adds cloud-identity collectors for GCP and Azure access enumeration in addition to secrets collection. StepSecurity reported targeting of GitHub Actions secrets, AWS, GCP, Azure, Kubernetes, HashiCorp Vault, npm tokens, and CircleCI tokens, plus explicit Harden-Runner evasion attempts.
  • SafeDep's June 1 technical analysis adds a concrete three-wave publication and payload shape: 64 malicious versions across 32 @redhat-cloud-services packages, first- and second-wave versions later unpublished but third-wave versions still the live latest for affected packages at the time of publication, a 4.3 MB index.js that ROT-9 decodes a loader, AES-128-GCM decrypts a 634 KB Bun script, downloads Bun 1.3.13 from GitHub Releases, and uses runtime paths such as /tmp/p<random>.js, /tmp/b-<random>/bun, and /tmp/kitty-<random>. SafeDep also reported Docker-socket escape attempts, Sigstore signing, .github/workflows/codeql.yml / Run Copilot repository persistence, and AI-agent / IDE persistence via .claude/settings.json and .vscode/tasks.json.
  • StepSecurity's runtime analysis adds a high-signal runner-memory theft path for Miasma: the preinstall payload located the GitHub Actions Runner.Worker process, used ACTIONS_RUNTIME_TOKEN to identify variables marked isSecret: true, and then attempted to read /proc/<pid>/mem to recover masked secrets directly from runner memory. StepSecurity also reported npm worming through stolen npm tokens and npm's bypass_2fa publish parameter, plus GitHub Contents API fallback exfiltration that routes stolen data through api.github.com commits in victim-controlled repositories.
  • JFrog's Miasma analysis, using @redhat-cloud-services/types@3.6.1 as the sample, reinforces that even type-only or generated-client packages can become install-time execution vehicles. JFrog lists network and host pivots including a configured https://api.anthropic.com/v1/api destination that returned a normal Anthropic 404 on plain GET and may be camouflage, GitHub commit-search C2 at https://api.github.com/search/commits?q=firedalazer, transient Bun payload paths under /tmp, malicious package metadata SHA256 7069e28a5806db4ab0273639667d203f5e31b401d403af7e36d9f360c1f6d655, and obfuscated install-loader SHA256 b86c5ae9e95bd841a595440faa3eb6317441e746f241ae8fd641ab59ed1d1966.
  • OX Security's June 2 follow-up characterizes the Red Hat / Miasma sample as a six-stage loader that can loop back into additional malicious execution by retrieving GitHub commits containing the firedalazer marker. OX reported two repository-description variants, Miasma: The Spreading Blight and Miasma : The Spreading Blight, warning that exact-string hunts for only the first form can miss infections. OX also highlighted different public encryption keys from Telnyx / LiteLLM-era TeamPCP samples and heavier obfuscation/encryption as reasons to keep this wave's attribution separate from simple TeamPCP assumptions.
  • JFrog also reported a destructive-token-monitor marker in Miasma commits: IfYouInvalidateThisTokenItWillNukeTheComputerOfTheOwner. Treat this as a containment-order warning for variants with active persistence: isolate affected hosts and stop malicious processes before broad GitHub-token revocation when feasible.
  • Microsoft Threat Intelligence's June 2 Miasma analysis counted 32 maliciously modified @redhat-cloud-services packages across more than 90 versions and tied the initial publishing path to the upstream RedHatInsights/javascript-clients CI/CD pipeline. Microsoft emphasized that the payload dynamically downloaded Bun for Linux, macOS, and Windows; Linux CI/CD runners appeared to be the primary target, but developer systems were also exposed to SSH, CLI, browser, and wallet theft.
  • Microsoft described the Miasma chain as a ten-phase flow: npm preinstall execution, staged ROT / AES-128-GCM unpacking, environment gating, defense evasion, credential access, passwordless-sudo privilege escalation, token-monitor persistence, three C2 / exfiltration channels including GitHub infrastructure abuse, forged-provenance package republishing, and a destructive decoy-token tripwire. Microsoft reported the tripwire could invoke rm -rf ~/ or ~/Documents, so responders should treat the IfYouInvalidateThisTokenItWillNukeTheComputerOfTheOwner marker as an active-host risk indicator rather than only a taunt.
  • Treat Miasma attribution as TTP overlap, not proof of TeamPCP operation. Wiz says the behavior is consistent with TeamPCP / Mini Shai-Hulud tradecraft, but the public release of Mini Shai-Hulud tooling makes a copycat plausible. Defender triage should include the vendor package lists, GitHub repositories with the Miasma: The Spreading Blight description, malicious orphan branches such as oidc-*, chore/add-codeql-static-analysis persistence branches, the GCP query user-agent google-api-nodejs-client/7.0.0 gl-node/20.11.0 gccl/7.0.0, api.github.com Contents API commit exfiltration, and any developer or CI host that installed the affected @redhat-cloud-services versions.
  • StepSecurity's June 25 Leo Platform report shows the Miasma-style payload factory continuing into a narrower single-ecosystem compromise: 20 Leo Platform npm packages, including leo-logger@1.0.8, leo-sdk@6.0.19, leo-aws@2.0.4, serverless-leo@3.0.14, leo-cli@3.0.3, and solo-nav@1.0.1, were maliciously published within a three-second window at 2026-06-24T23:04:55Z. StepSecurity reported the same binding.gyp "Phantom Gyp" hook, ROT-N / AES-128-GCM / obfuscator.io layering, Bun v1.3.13 runtime download, Runner.Worker memory scraping, GitHub dead-drop exfiltration, npm bypass_2fa worming, workflow injection, and runner ALL=(ALL) NOPASSWD:ALL sudoers modification seen in Miasma. Socket's same-day follow-up expanded the wave to 23 npm packages by adding hexo-deployer-wrangler@1.0.4, hexo-shoka-swiper@0.1.10, and prism-silq@1.0.1 from publisher llxlr, connected the wave to RevokeAndItGoesKaboom GitHub commit-search/dead-drop tradecraft seen in adjacent GitHub Actions compromises, and reported a related github.com/verana-labs/verana-blockchain@v0.10.1-dev.20 Go source archive carrying .claude/ and .vscode/ hooks that execute the same Bun-staged payload when a trusted IDE or AI coding assistant opens the repository. Track this as likely same payload factory / Miasma-style activity, but keep actor attribution caveated because public tooling and copycat reuse remain plausible.
  • StepSecurity's June 26 Immobiliare Labs report adds a Backstage-plugin target set to the same Miasma / Phantom Gyp lineage: four npm packages (@immobiliarelabs/backstage-plugin-gitlab, @immobiliarelabs/backstage-plugin-gitlab-backend, @immobiliarelabs/backstage-plugin-ldap-auth, and @immobiliarelabs/backstage-plugin-ldap-auth-backend) had malicious patch releases inserted across supported major lines within about 30 seconds. StepSecurity's diff of @immobiliarelabs/backstage-plugin-gitlab@2.1.2 against clean 2.1.1 found a newly added 5 MB root index.js and binding.gyp; the payload used ROT-2, AES-128-GCM, obfuscator.io layering, Bun v1.3.13 staging, Runner.Worker /proc/<pid>/mem secret scraping, cloud / package-registry / password-manager credential collection, AI-assistant persistence through files such as .github/copilot-instructions.md, and apparent npm/PyPI worming functions. This reinforces the current defensive priority: treat newly published patch releases in internal-developer-platform dependencies as high-risk until cooldown, tarball diffing, and install-time execution review have completed.
  • JFrog reported a TeamPCP-linked / possible-copycat compromise of the legitimate xinference PyPI release line: versions 2.6.0, 2.6.1, and 2.6.2 executed from xinference/__init__.py on import, spawned a detached Python payload, collected developer/cloud/Kubernetes secrets, and exfiltrated love.tar.gz to whereisitat[.]lucyatemysuperbox[.]space with header X-QT-SR: 14. TeamPCP publicly denied involvement, so keep attribution caveated.
  • JFrog reported more than 170 npm packages and 2 PyPI packages affected in its earlier analysis window, with npm payloads using malicious preinstall loaders and PyPI payloads using import-time downloaders.
  • JFrog's May 19 follow-up counted the AntV wave as 325 legitimate npm packages after identifying @cap-js/openapi@1.4.1; the added package used a cleaner-looking optionalDependencies reference to github:cap-js/openapi#d78c25443ec4a0d7f0a85776461f3b1163132537 and delivered the Shai-Hulud payload from fork-resolvable GitHub content rather than embedding malicious code directly in the tarball.
  • Socket reported continuing package findings across npm and PyPI ecosystems, including OpenSearch, Mistral AI, Guardrails AI, Squawk, and other artifacts in related coverage. In its May 12 update, Socket named @opensearch-project/opensearch npm versions 3.5.3, 3.6.2, 3.7.0, and 3.8.0, PyPI mistralai@2.4.6, PyPI guardrails-ai@0.10.1, and additional @squawk/* packages such as @squawk/mcp@0.9.5, @squawk/weather@0.5.10, and @squawk/flightplan@0.5.6.
  • Socket's guardrails-ai@0.10.1 analysis is operationally important because it used Python import-time execution rather than npm lifecycle execution: on Linux it downloaded https://git-tanstack[.]com/transformers.pyz to /tmp/transformers.pyz and executed it with python3 without integrity verification. Socket reported the git-tanstack[.]com host displayed a TeamPCP-signed taunt while the infrastructure was live, giving defenders another infrastructure and import-time-loader pivot for the May 11 wave.
  • Socket also reported a separate February 2026 Shai-Hulud-style SANDWORM_MODE cluster that used typosquatted npm packages and a malicious GitHub Action to target CI secrets and AI coding toolchains; track it separately because the delivery model and MCP prompt-injection tradecraft differ from the May Mini Shai-Hulud waves.
  • StepSecurity, Snyk, and Unit 42 reported an AntV-centered wave involving the atool maintainer account, timeago.js, echarts-for-react, and many @antv/* visualization packages. StepSecurity described a two-wave May 19 publish pattern: first using a preinstall hook that invoked Bun, then adding Bun as an explicit dependency to improve delivery reliability.
  • Unit 42 counted approximately 639 malicious package versions across 323 unique packages in about one hour, calling it the largest single-hour package count of any Shai-Hulud wave observed in its reporting.
  • Socket later summarized the same AntV burst as 639 malicious versions across 323 unique packages after the atool npm maintainer account was seized, and said npm responded on May 19 by invalidating all granular access tokens with write access that bypass two-factor authentication. Socket's campaign rollup counted 1,055 Mini Shai-Hulud package versions across 502 unique packages at that point: 1,048 npm versions across 498 packages, six PyPI entries across three packages, and one Composer package-version entry.
  • SafeDep's May 19 AntV / atool analysis adds several concrete detection pivots for the same wave: it counted 637 malicious versions across 317 packages in a roughly 22-minute automated burst, reported payload SHA256 a68dd1e6a6e35ec3771e1f94fe796f55dfe65a2b94560516ff4ac189390dfa1c, and identified orphan/imposter antvis/G2 commits 1916faa365f2788b6e193514872d51a242876569, 7cb42f57561c321ecb09b4552802ae0ac55b3a7a, and dc3d62a2181beb9f326952a2d212900c94f2e13d as redundant payload-delivery paths via @antv/setup optional dependencies. SafeDep also described dual exfiltration through GitHub dead-drop repositories and t.m-kosche[.]com OpenTelemetry-looking HTTPS posts, Docker-socket host escape attempts, local Node.js project propagation, .github/workflows/codeql.yml persistence named Run Copilot, AI-assistant SessionStart hooks for Claude Code / Codex, VS Code runOn: folderOpen tasks, kitty-monitor GitHub commit-search C2 using the firedalazer keyword, and a gh-token-monitor daemon that repeatedly polled stolen GitHub tokens.
  • Wiz's May 19 AntV writeup adds a compact TeamPCP-attribution and persistence pivot set for the same wave: moderate-confidence TeamPCP attribution based on infrastructure, malware behavior, and operational overlap; backup C2 m-kosche[.]com / 185.95.159.32; GitHub commit search for api.github.com/search/commits?q=firedalazer; persistent backdoor path ~/.local/share/kitty/cat.py; macOS LaunchAgent ~/Library/LaunchAgents/com.user.kitty-monitor.plist; Linux user service ~/.config/systemd/user/kitty-monitor.service; and backdoor SHA256 fb5c97557230a27460fdab01fafcfabeaa49590bafd5b6ef30501aa9e0a51142.
  • Treat that npm-wide reset as an incident-response interruption, not a root-cause fix: Socket noted that it burns already stolen bypass-2FA tokens but does not address workflow-level publication paths such as TanStack-style OIDC extraction, cache poisoning, or compromised trusted-publishing pipelines.
  • GitHub's May 22 npm changelog made staged publishing generally available in npm CLI 11.15.0 or newer. The feature lets CI upload a prebuilt tarball into a stage queue, then requires a human maintainer to pass a 2FA challenge before the package becomes installable; it can be paired with OIDC trusted publishing and configured so a workflow may only run npm stage publish, not direct npm publish. This is a meaningful mitigation for unattended CI publication, but not a substitute for cleaning compromised workflows, runners, caches, and maintainer endpoints.
  • The same npm release added install-source allowlist controls: --allow-file, --allow-remote, and --allow-directory, complementing --allow-git. These flags matter for Mini Shai-Hulud-style defense because several waves used non-registry GitHub/orphan-commit or remote runtime paths instead of obvious malicious code embedded directly in the registry tarball.
  • StepSecurity, JFrog, and Unit 42 reported that AntV-wave payloads read GitHub Actions runner process memory to recover masked CI/CD secrets, harvested developer/cloud/Kubernetes/Vault/crypto-tool paths, queried local password-manager CLIs including 1Password, Bitwarden, pass, and gopass, exfiltrated through a GitHub dead-drop and t.m-kosche[.]com, and created public Dune/Shai-Hulud-themed repositories from stolen tokens. JFrog also reported that the npm payload logic could request GitHub Actions OIDC material, exchange it for npm trusted-publishing credentials, and create Sigstore provenance, reinforcing that valid provenance can be produced by a compromised workflow.
  • StepSecurity and Snyk reported malicious durabletask PyPI versions 1.4.1, 1.4.2, and 1.4.3 in Microsoft's official Durable Task Python SDK. Unlike the TanStack trusted-publishing chain, these uploads reportedly bypassed the GitHub release workflow and used real PyPI publishing credentials.
  • The durabletask payload was reported as a Linux-focused Python zipapp (rope.pyz) that harvested AWS, Azure, GCP, Kubernetes, password-manager, and developer-tool secrets, used redundant exfiltration paths, installed fake systemd persistence, attempted lateral movement via AWS SSM and Kubernetes kubectl exec, skipped Russian-locale systems, and used TeamPCP-linked infrastructure (t.m-kosche[.]com). JFrog reported AWS SSM propagation state under /tmp/.rope_state/ssm_instances.json, an SSM marker at ~/.cache/.sys-update-check, Kubernetes propagation marker ~/.cache/.sys-update-check-k8s, and attempts to propagate to up to five non-Windows SSM instances or Kubernetes pods where permissions allowed. Wiz additionally reported primary C2 at check.git-service[.]com, downloaded payload paths such as /tmp/managed.pyz and /tmp/rope-*.pyz, and infection markers ~/.cache/.sys-update-check and ~/.cache/.sys-update-check-k8s.
  • StepSecurity's June 6 Azure/durabletask follow-up reported that the same contributor account later pushed malicious editor / AI-agent configuration files into the GitHub repository, and that GitHub disabled 73 Microsoft repositories across four organizations in a 105-second enforcement window. The disabled set included Azure/functions-action, causing workflows pinned only to Azure/functions-action@v1 to fail while the repository was unavailable. Treat this as another Mini Shai-Hulud / Miasma blast-radius pattern: repository-level AI-tool persistence can trigger platform enforcement and availability impact for downstream CI/CD consumers even without a new registry package release.
  • OX Security's June 4 Miasma / binding.gyp update adds another hunting nuance for the same wave: it measured 57 affected packages, 152,376 accumulated weekly downloads, 647,204 accumulated monthly downloads, and more than 118 GitHub repositories with stolen credentials, while reporting the repository-description marker as Miasma - The Spreading Blight with an en dash. Include that punctuation variant in GitHub dead-drop searches alongside the colon, spaced-colon, and hyphen variants already reported by other vendors.
  • Socket and StepSecurity's June 7-8 reporting describes a Hades PyPI branch of the same Mini Shai-Hulud / Miasma lineage. Socket initially counted 37 malicious PyPI wheels across 19 packages and later tracked 60 PyPI artifacts across 37 packages, mixed across real bioinformatics / graph-learning packages (dynamo-release, spateo-release, coolbox, ufish, napari-ufish, embiggen, ensmallen, gpsea, pyphetools, ppkt2synergy) plus MCP / LangChain / OpenAI-themed bait such as langchain-core-mcp, instructor-mcp, openai-mcp, ray-mcp-server, and tiktoken-mcp. Treat the package list as vendor-maintained and fast-moving rather than reproducing it exhaustively here.
  • Hades changes the Python execution lane. The first PyPI wave used *-setup.pth startup hooks with a bundled _index.js; Socket's follow-up reported native-extension import-time execution through trojanized .abi3.so files and a langchain-core-mcp loader that searches sys.path for _index.js before running it with Bun. This evades hunts that only diff Python source or require the JavaScript payload to live in the same wheel as the startup hook.
  • The Hades payload keeps the Mini Shai-Hulud credential-theft target set: GitHub and runner secrets, npm / PyPI / RubyGems / JFrog / CircleCI publishing material, AWS / GCP / Azure / Kubernetes / Vault secrets, SSH keys, Docker config, shell histories, .env, .npmrc, .pypirc, Claude / MCP configuration, and other developer-machine credentials. StepSecurity reported cross-platform runner-memory scraping, SSH/SCP lateral movement using ~/.ssh/known_hosts and ~/.ssh/config, PyPI/npm OIDC and provenance abuse attempts, AI-assistant / IDE workspace backdooring, and a gh-token-monitor wiper-deterrent service that can run destructive deletion when a stolen GitHub token starts returning 4xx responses.
  • Hades also changes GitHub hunt vocabulary. Socket reported dead-drop repositories with description Hades - The End for the Damned and generated names using terms such as stygian, tartarean, cerberus, charon, styx, lethe, thanatos, and persephone. StepSecurity reported three GitHub-search C2 / exfiltration markers: DontRevokeOrItGoesBoom for encrypted token commits, TheBeautifulSnadsOfTime for signed JavaScript eval payloads, and firedalazer for a Python updater.py polling channel.
  • StepSecurity's June 11 developer-machine detection note frames the current Miasma/Hades defensive gap as open-time and import-time execution, not just install-time package review. It highlights four quiet project-tree artifacts defenders should inventory on endpoints and recently cloned repositories: a 157-byte binding.gyp used by Miasma's "Phantom Gyp" native-build path, injected Python __init__.py import hooks for Hades, .vscode/tasks.json folder-open tasks, and .claude/setup.mjs / Claude hook launchers. These files may not appear in dependency manifests or lockfile diffs, so endpoint and repository hunts need file-content rules in addition to registry/package-version checks.
  • JFrog's June 15 prompt-injection analysis adds a different Hades/Miasma-era evasion path: a Shai-Hulud sample placed plain-English text before the obfuscated payload to trigger AI scanner safety refusals rather than to win a false "clean" verdict. JFrog reported that some guarded model/API setups refused or suppressed malware-analysis output before returning the suspicious eval / obfuscation findings, while nonguarded or differently wrapped model paths still identified the malicious code. Treat an AI scanner refusal as a null or suspicious result, not as a pass, and send the artifact through static rules, sandboxing, or human review.
  • Grafana Labs publicly stated that the TanStack/Mini Shai-Hulud incident led to unauthorized access to its GitHub environment and source-code download after one impacted workflow token was missed during rotation. Grafana reported no evidence of production-system or Grafana Cloud compromise and said its codebase was downloaded but not altered.
  • CrowdStrike's 2026 Threat Hunting Report adds cross-campaign field context without resolving TeamPCP attribution. CrowdStrike says ALTERED SPIDER compromised more than 300 software dependencies in one day, harvested credentials, and pivoted into cloud environments; it separately attributes the March Axios / ZshBucket compromise and June Mastra / malicious-dependency incident to DPRK-nexus STARDUST CHOLLIMA. CrowdStrike counted npm in 87% of identified software-registry threats during the first half of 2026. Keep those vendor labels scoped to CrowdStrike's reporting rather than treating ALTERED SPIDER as a confirmed TeamPCP alias, and use the 87% figure as telemetry from one provider, not a registry-wide prevalence estimate.
  • Wiz's H1 2026 review adds provider-side field context for the campaign's long tail. Significant incidents highlighted to Wiz customers rose 60% from H2 2025; supply-chain incidents more than doubled and grew from about 10% to 25% of the highlighted set. Wiz says TeamPCP moved through hundreds of organizations, listed roughly 4,000 stolen private repositories for sale at one point, and produced credentials that were reused weeks later in hands-on data-theft intrusions by groups Wiz believes were different operators. Keep those figures scoped to Wiz telemetry and preserve the distinction between TeamPCP's theft, later third-party credential reuse, and copycats using the May 12 open-source release.

May 2026: adjacent GitHub Actions and IDE-extension lanes

  • StepSecurity reported compromised actions-cool/issues-helper and actions-cool/maintain-one-comment GitHub Actions where all release tags were moved to imposter commits. The malicious action downloaded Bun, read Runner.Worker memory for decrypted workflow secrets, and exfiltrated to t.m-kosche[.]com, matching infrastructure and runner-memory-theft motifs seen in the broader Mini Shai-Hulud cluster.
  • StepSecurity reported a compromised Nx Console VS Code extension (nrwl.angular-console 18.95.0) that fetched an obfuscated payload from an orphan commit in the official nrwl/nx repository. This is not the same registry lane as npm/PyPI worming, but it targets the same developer-trust boundary.
  • GitHub publicly confirmed a May 18 employee-device compromise involving a poisoned third-party VS Code extension and linked to the Nx Console security advisory; GitHub said the activity involved exfiltration of GitHub-internal repositories only and that the attacker's roughly 3,800-repository claim was directionally consistent with its investigation.
  • See also: actions-cool GitHub Actions tag compromise and Nx Console VS Code extension compromise.

Tradecraft map

Initial access / publication path

  • pull_request_target or similar privileged workflow footguns that run fork-controlled code in a privileged repo context.
  • GitHub Actions cache poisoning or runner-state poisoning that survives until a legitimate release workflow executes.
  • OIDC/trusted-publishing token extraction from runner memory, then exchange for short-lived npm publishing credentials.
  • Compromised maintainer/package publisher accounts in some waves.

Execution and payload staging

  • npm lifecycle hooks such as preinstall.
  • Bun runtime download/execution to run large JavaScript payloads.
  • PyPI import-time loader/downloader behavior in related Python packages.
  • Composer plugin install/update hooks and mutable Packagist tag metadata in PHP ecosystem compromises.
  • Heavy obfuscation and embedded encrypted payload sections.

Credential harvesting

  • GitHub PAT/OAuth tokens and Actions runtime secret material.
  • npm tokens and trusted-publishing exchange material.
  • AWS, Azure, GCP, Kubernetes, Docker, Vault, Terraform, SSH, Git, shell history, .npmrc, cloud config, and generic API secrets.
  • Kubernetes API enumeration where service-account permissions allow it.
  • Browser/password-store collection reported in later variants.
  • Credential-location expansion (GitGuardian, Aug-Sep 2026): a recent Mini Shai-Hulud wave (the keyv incident, hitting keyv and 800+ npm packages) broadened the worm's file-system secret collector from 189 to 469 hardcoded locations. GitGuardian's per-OS breakdown of the collector: Linux 89→290, Windows 12→50, macOS 88→129. The added paths are concentrated in developer environments, CI/CD tooling, cloud configuration, crypto-wallet locations, and AI tool/agent configuration (coding-agent settings, CLI caches, IDE secrets). The durable defender read: the worm no longer breaks trust relationships, it harvests the credentials that already make them work (publishing tokens, cloud keys, CI runner identities), so the highest-value exposure is any standing, long-lived package-publishing credential on a reachable host.

Exfiltration and propagation

  • Encrypted exfiltration via attacker-controlled infrastructure.
  • GitHub fallback/dead-drop repositories created in victim accounts.
  • Repo naming/description patterns reported by vendors, including Dune/Shai-Hulud themed descriptions and configuration-storage masquerades.
  • Automated enumeration of packages the victim can publish, tarball modification, version bumping, metadata injection, and republishing.
  • Repository poisoning through .claude/, .codex/, and .vscode/ files in variants that try to reach AI coding agents and IDE automation. JFrog reported SessionStart hook injection for Claude Code/Codex settings and a VS Code folderOpen task path in the May 19 wave.
  • GitHub commit-search C2/persistence: JFrog reported kitty-monitor, which searched GitHub commits for signed command markers such as firedalazer; the @cap-js/openapi variant used separate markers including thebeautifulsnadsoftime and thebeautifulmarchoftime.
  • GitHub Actions tag retargeting as an adjacent lane: trusted action tags can be moved to imposter commits, allowing malicious runtime code to read runner memory and steal secrets.
  • IDE-extension compromise as an adjacent lane: poisoned VS Code extensions can reach developer endpoints even when package lockfiles and build dependencies are clean.

Persistence / destructive behavior

  • Claude Code hooks and VS Code task automation reported as persistence or re-execution paths.
  • Background daemon behavior and dead-man-switch style deletion/wiping behavior reported in later Shai-Hulud/Mini Shai-Hulud analysis.
  • Developer endpoints and CI runners should be treated as compromised hosts, not just as places where a bad dependency was installed.

Defender heuristics

Exposure triage

  • Search dependency lockfiles, package-manager caches, CI logs, and artifact repositories for affected package names/versions from vendor advisories.
  • Treat any install of affected versions in CI or on a developer machine as credential exposure.
  • Prioritize environments with npm publishing permissions, GitHub org/admin tokens, cloud deployment credentials, Kubernetes service-account access, or Vault access.

GitHub and CI hunting

  • Look for pull_request_target workflows that check out or execute fork-controlled code.
  • Review caches restored by release workflows, especially caches writable by pull-request jobs.
  • Hunt workflow logs for unexpected Bun downloads, large obfuscated JavaScript payloads, preinstall execution, runner memory scraping, or token/OIDC environment access.
  • Search for unexpected repositories created by maintainers/bots with Shai-Hulud/Dune/config-storage descriptions or encrypted blobs; Unit 42 notes that later variants can use both GitHub dead-drop repositories and telemetry-looking HTTPS exfiltration to t.m-kosche[.]com. Socket reported AntV-wave dead-drop repositories using reversed markers such as niagA oG eW ereH :duluH-iahS / niaga og ew ereh :duluh-iahs, Dune-themed names like sayyadina-stillsuit-852, and results/results-*.json paths.
  • For the June 2026 Miasma wave, search GitHub for repositories with description Miasma: The Spreading Blight, review @redhat-cloud-services/* package installs against the Wiz and StepSecurity affected-version lists, and hunt for GCP API traffic using google-api-nodejs-client/7.0.0 gl-node/20.11.0 gccl/7.0.0 from developer or CI environments.
  • Include OX's alternate Miasma repository-description spacing, Miasma : The Spreading Blight, and GitHub commit-search C2 marker firedalazer in hunts; avoid relying on a single literal description string.
  • For the Hades PyPI branch, search for Hades - The End for the Damned, Underworld-themed repository names, DontRevokeOrItGoesBoom, TheBeautifulSnadsOfTime, thebeautifulmarchoftime, firedalazer, *-setup.pth files paired with _index.js, .abi3.so extensions paired with _index.js, langchain_core-setup.pth, /tmp/.sshu-setup.js, and unexpected Bun execution from Python interpreter startup or native-extension import paths.
  • Audit newly added .claude/, .codex/, and .vscode/ files, especially settings.json, tasks.json, setup.mjs, copied payload scripts, Claude Code/Codex SessionStart hooks, and VS Code folderOpen tasks. Include source-tree scans for standalone binding.gyp native-build triggers and suspicious Python package __init__.py import hooks, because StepSecurity's June 11 Miasma/Hades note emphasizes that these execution lanes can bypass manifest- and lockfile-centered review.
  • Search for SafeDep-reported AntV indicators where applicable: payload hash a68dd1e6a6e35ec3771e1f94fe796f55dfe65a2b94560516ff4ac189390dfa1c, @antv/setup GitHub optional-dependency references to orphan antvis/G2 commits, .github/workflows/codeql.yml workflows named Run Copilot, ~/.local/bin/gh-token-monitor.sh, ~/.local/share/kitty/cat.py, kitty-monitor.service, com.user.kitty-monitor.plist, /var/tmp/.gh_update_state, and GitHub commits containing firedalazer command markers.
  • For the Leo Platform wave, search package-locks, CI caches, artifact repositories, and developer endpoints for the 20 StepSecurity-listed package/version pairs published at 2026-06-24T23:04:55Z; prioritize leo-logger@1.0.8 and leo-sdk@6.0.19 where StepSecurity performed deeper analysis, then expand to the full Leo package set.
  • Include Socket's Leo / Miasma follow-up pivots: hexo-deployer-wrangler@1.0.4, hexo-shoka-swiper@0.1.10, prism-silq@1.0.1, github.com/verana-labs/verana-blockchain@v0.10.1-dev.20, .claude/index.js, .claude/setup.mjs, .vscode/setup.mjs, .claude/settings.json, .vscode/tasks.json, RevokeAndItGoesKaboom, and Alright Lets See If This Works.
  • Search GitHub audit logs and repositories for suspicious workflow commits matching reported Mini Shai-Hulud patterns such as branch chore/add-codeql-static-analysis, commit message fix: ci, and unexpected .github/workflows/codeql.yml content labelled Run Copilot.
  • Hunt for post-theft GitHub activity such as many git.clone events, malicious workflow pull requests, workflow-log deletion, or activity from VPN / VPS ASNs not normally used by maintainers.
  • In AWS, alert when EC2 instance-role credentials are used from external or non-instance addresses. For CI identities, correlate IMDS access and unusual runner egress with CreateUser, AttachUserPolicy, CreateAccessKey, broad Secrets Manager enumeration, database security-group changes, Redshift Data API activity, SSM SendCommand, and SES identity-verification calls. Hunt the FortiGuard case pivots cloudops-monitor, exfil-s3-write, exfil-s3-full, exfil* STS sessions, 185.204.1[.]225, and 89.22.231[.]63:8080, validating current ownership before blocking public infrastructure.
  • Inventory IDE extensions on developer machines; treat a malicious editor extension as an endpoint compromise capable of reading source, secrets, shell history, and authenticated GitHub sessions.

Package and registry hunting

  • Diff newly published package tarballs against prior clean versions.
  • Flag new lifecycle hooks, new Bun/runtime downloaders, large minified/obfuscated payload files, or sudden patch releases from unusual automation.
  • For Composer/Packagist, flag packages that unexpectedly become composer-plugin packages, add composer-plugin-api, introduce post-install-cmd / post-update-cmd execution paths, or move an existing version tag to a new commit.
  • Do not trust provenance alone; correlate attestations with clean workflow inputs, clean cache state, and expected release commits.
  • Add release-age/cooldown controls for package ingestion when operationally possible.
  • If an LLM- or agent-based scanner refuses to analyze a package, quarantine/escalate the result instead of treating it as clean. Keep deterministic detections for lifecycle hooks, root-level obfuscated eval, unexpected Bun/runtime fetches, type-only packages with executable payloads, and prompt-like text that attempts to manipulate the analyzer.
  • For npm maintainers, prefer staged publishing for high-blast-radius packages: CI should stage the tarball and a separate human 2FA approval should release it from a trusted device. Where available, configure trusted-publisher bindings as stage-only so direct CI npm publish is rejected.
  • For npm and PyPI trusted publishing, bind publishers to GitHub Environments with deployment-branch rules where possible; workflow filename matching alone can authorize a throwaway-branch workflow if an attacker has repository write access.
  • In npm consumers, consider setting non-registry install sources to deny-by-default (--allow-git=none, --allow-remote=none, --allow-file=none, --allow-directory=none) and only allow exceptions after review; this specifically reduces exposure to GitHub/orphan-commit and remote-tarball dependency paths used in recent supply-chain waves.
  • After registry-wide token resets, explicitly inventory and replace automation tokens that stopped working, but do not assume new tokens are safe until affected runners, developer endpoints, caches, and release workflows have been cleaned.

Containment

  • Stop affected workflows, isolate affected hosts, and package publication paths before rotating secrets if persistence or active exfiltration may still be running; Unit 42 specifically warns that some May 11 payloads used token-revocation-triggered destructive behavior while the daemon was active.
  • Remove malicious packages and poisoned repo files, then rotate all reachable credentials: GitHub, npm, cloud, Kubernetes, Vault, SSH, Docker, CI, and any app secrets present on the host.
  • Invalidate GitHub Actions caches and rebuild release infrastructure from known-clean commits.
  • Prefer short-lived scoped credentials, protected environments, least-privilege OIDC subjects, pinned action SHAs, and separate untrusted PR workflows from release workflows.

Monitoring notes

  • Fourteen months on, the Sep-2025 worm's own paperwork is still producing false positives: MAL-2025-47143 @ctrl/transmission carries an open introduced: "0" range and scanners that evaluate ONLY ranges now report CLEAN releases as malicious (queue filing ossf/malicious-packages #1803, drakonkat, Nov-18 11:55:26Z, OPEN; added by the 327th sweep): the row for the original September-2025 @ctrl/* worm victim lists versions ["7.3.1"] with no bounded SEMVER range, unlike its siblings (MAL-2025-47131 @ctrl/deluge and MAL-2025-47138 @ctrl/qbittorrent were narrowed in #1003 with introduced/last_affected next to the same versions list). Consequence in the wild, per the filing and reproduced by the reporter's JetBrains scan: the dependency check ("Results powered by Mend.io") flags @ctrl/transmission 7.4.0 and 8.0.1 - releases published AFTER the takedown, 8.0.0/8.0.1 from GitHub Actions Oct-5 with SLSA provenance - as malicious, while the sibling @ctrl/* packages scan clean; GHSA-8v7r-ghw6-g8gq correctly lists = 7.3.1, and npm removed 7.3.1 on 2025-09-16 (maintainer shipped 7.4.0 the next morning). THIS WIKI verified ≈12:4xZ: OSV row modified frozen 2025-09-17T06:23:36Z (untouched for ≈14.5 months), affected[0] versions ['7.3.1'] + ranges None, aliases ['GHSA-8v7r-ghw6-g8gq']; sibling MAL-2025-47131 verified carrying its bounded SEMVER introduced 7.2.1 / last_affected 7.2.2; npm @ctrl/transmission latest 8.0.1, 7.3.1 absent from roster, 7.3.1 tarball 404. AND THE FILING'S FLAGGED SIBLINGS SHARE THE BUG - verified same sweep: MAL-2025-47200 ngx-trend 8.0.1, MAL-2025-47272 @teselagen/bounce-loader 0.3.16/0.3.17, MAL-2025-47274 @teselagen/liquibase-tools 0.4.1 are ALL versions-only with ranges None, modified frozen 2025-09-17 = a FIVE-NAME open-range false-positive cohort from the Sep-2025 flush still shipping fourteen months later. DURABLE READS: (1) advisory paper QUALITY is a defensive surface - a versions-only row with an unbounded ranges entry is a scanner bug wearing a malware row's clothes; before escalating a critical hit on a package's CURRENT release, check whether the poisoned rung was actually that version and what affected[0].ranges says; (2) a critical on a release that post-dates the takedown deserves a shape check first, deletion of the poisoned tarball (here 7.3.1 404 since Sep-16-2025) plus provenance attestations on the later releases is the exoneration evidence; (3) the merge-gate clock predicts the fix: #1800's merge 03:02:10Z produced the 20062 row-edit 03:15:04Z ≈13 m later, so watch whether #1803 merging lands a bounded range + modified bump on 47143 (and whether the three un-flagged siblings get widened by the same pass or stay frozen = per-row remediation, not per-family). No new worm activity - this is AFTERMATH PAPER on the Sep-2025 original; quote the date.
  • The worm repackages itself onto RubyGems, strutil-core (added by the 238th sweep, Nov 10, 2026): RubyGems account jamnajiea published a clean 1.0.0 on Nov-1, then the malicious 1.0.1 under two hours later; flagged 70 s after release and removed ≈5 h later (RubyGems now answers could not be found for both versions; Socket says no victims observed). OSV MAL-2026-19054 (published Nov-1, forensics WRITTEN Nov-10 00:00Z, nine days after the row existed) describes the testmetest/Miasma-family rebuild: the native-extension install step downloads Bun from Bun's own GitHub releases into /tmp/.b_<pid>/ and runs ext/strutil-core/index.js, which read 51 credential stores (AWS/Azure/GCP/Vault; GitHub CLI, GitLab CLI, git creds, SSH; npm/PyPI/RubyGems/NuGet/Composer/Yarn; Docker, K8s, Claude Code and Codex logins, DB clients, crypto wallets, Terraform/Databricks/Hugging Face/Vercel/Kaggle) and then used them live, AWS STS/Secrets Manager/SSM Parameter Store across 18 regions and ECR in 11, Azure ARM + Microsoft Graph, GCP metadata + Resource Manager, GitHub/GitLab/GHCR APIs. Persistence: systemd user service ~/.config/systemd/user/pkg-runtime.service + ~/.local/share/runtimes/sync.py + ~/.config/index.js, and it rewrote ~/.cursor/mcp.json (AI-tool config poisoning as a gem). Exfil server paleasbest.top is the same server the pystr-codec PyPI worm used (MAL-2026-18432, kam193, Nov-1), per the OSV's own analysis this is the same worm repackaged across registries; Socket traces the module to the deobfuscated June 2026 Miasma payload (33 of 36 non-empty installer lines verbatim). Durable reads: (1) the Shai-Hulud-derivative worm kit is now cross-registry by default, npm/PyPI/RubyGems republish routines ship in one payload; (2) the AI-tool credential set (.claude/.credentials.json, .codex/auth.json, Cursor MCP config) is a first-class target in every generation; (3) the 70 s flag / 5 h removal clock shows RubyGems' abuse lane now runs at npm-machine speeds, the gap that mattered was the ≈2 h window between the clean and malicious publishes, which no post-hoc takedown covers. Hunt: retro-scan developer hosts and CI images for /tmp/.b_* remnants, pkg-runtime.service, sync.py, ~/.config/index.js, and ~/.cursor/mcp.json edits; watch paleasbest.top and the jamnajiea-class account grammar; enumerate gem pushes in the Nov 1-10 window for siblings. Sources: Socket via GBHackers (Nov 10) https://gbhackers.com/ruby-gem-discovered/amp; OSV MAL-2026-19054 https://api.osv.dev/v1/vulns/MAL-2026-19054; MAL-2026-18432 https://api.osv.dev/v1/vulns/MAL-2026-18432; Socket official post (Nov 11 capture, added by the 241st sweep) https://socket.dev/blog/shai-hulud-descendant-malicious-ruby-gem
  • The official Socket post (captured Nov-11 by the 241st sweep) moves the detail lane with four durable additions: (1) REVOKED-TOKEN HOSTAGE MONITOR: when a stolen GitHub token is accepted the worm installs a background monitor, and if the token later returns 4xx (revoked) it deletes the user's home directory = the hostage dead-man grammar (binding-gyp gh-token-monitor, tensorlake lineage) lands on the RubyGems generation, THIRD on-wiki class member — remediation ORDER stands: isolate hosts and preserve evidence BEFORE revoking tokens; (2) COMMIT-MESSAGE-AS-C2: a Python backdoor downloads and runs ANY script the threat actor announces in a signed GitHub commit message = the C2 channel lives in a field no package-side scanner reads; hunt for repos whose commit-message entropy/length breaks cadence and for fetch-run-pipe patterns anchored on commit metadata; (3) REPO-PLANTED AI RE-ARM: with stolen git credentials it commits CLAUDE.md, AGENTS.md, .cursor/mcp.json, and a .vscode/tasks.json task INTO repositories — opening an affected repo in an AI coding tool re-runs the worm depending on trust settings = the config-auto-execution law's repository-resident variant, the infection persists in SOURCE CONTROL not just local config; (4) VICTIM-FLOW COUNTER-FORENSICS: the ruby-digger metadata tracker shows 89 downloads accrued to the CLEAN 1.0.0 (mirrors/scanners) and 0 to malicious 1.0.1 in the 45 s before flagging = no victim flow touched the bad rung pre-removal, the downloads lane again remembers what the registry forgets. StepSecurity adds: runner-memory scrape in GitHub Actions (tj-actions grammar) and an UNSUCCESSFUL attempt to disable Harden-Runner (the anti-EDR tell as a hunt shape).
  • The worm never stopped: Metano's live-fleet measurement (reported ~Nov 2-3, 2026; disclosed to GitHub Nov 2, added to this page by the 237th sweep Nov 6): the Mini Shai-Hulud dead-drop infrastructure has been ACTIVE CONTINUOUSLY since Oct 21, 990 public dead-drop repositories holding encrypted stolen secrets created since Oct 21 (newest observed Nov 3 07:45 UTC), on top of ≥1,500 more created Sep 17-Oct 20 and since deleted (≥2,500 total). Estimated 100-200 infected machines (grouped by upload file size), ~60 of the groups re-appearing across multiple days = persistent infections, ~2 new repeat senders/day = a slow trickle of NEW infections, not a burst. The token economy is the headline: uploads almost always ride a token STOLEN FROM ANOTHER DEVELOPER, passed between infected computers, when one token dies, another infected host notices and shares its own; nobody manages it by hand. The worm also adds a secret-stealing GitHub Actions workflow on a fake Dependabot branch and deletes it within ≈10 seconds (too fast for any repo owner to notice; runs only with the host's own workflow-scoped token, so a push = proof the worm ran on that developer's machine, the shared-token owners are victims too). Almost NO poisoned packages back this wave (the only reported one remains tensorlake, Oct 8), the campaign has decoupled from npm publishing entirely and now lives purely on the credential fleet; the code still carries republish capability and can re-arm the package lane the moment a 2FA-exempt npm token appears. The May 2026 worm is ALSO still live: present in 139 forks of a popular open-source AI chat app via the hijacked actions-cool sync workflow (137 still carrying it, 108 on main), one fork owner's machine re-running it Nov 1. Defender reads: dead-drop repos are a live census, a public-GitHub watch for the worm's repo-description fingerprints is a fleet telemetry source; the shared-token relay means revoking YOUR tokens does not stop the fleet, and stolen tokens circulate as infrastructure, treat any credential that ever touched an infected host as burned for months; and the 10-second workflow delete makes audit-log (not repo-state) monitoring mandatory for Actions workflow events on high-value orgs. Source: https://metano.ai/resources/mini-shai-hulud-is-back
  • ATT&CK baselining + hostage-token literal + generational naming table (Oct 10, 2026 one-hundred-and-thirty-seventh sweep, priority-watch durable adds): (1) MITRE ATT&CK now carries software S9043 "Mini Shai-Hulud", credential stealer + self-replicating supply-chain worm with a targeted wiper component, mapped to group G1056 "TeamPCP", techniques including T1003.007 (proc-filesystem runner-memory token scraping) and T1195.001. The campaign pair is now in the BASELINE matrix: commercial detection engines and CTI platforms inherit the coverage by default, so teams that deferred detection "until attribution firms agree" have lost that excuse. Attribution discipline per this page's standing rule stays: S9043→G1056 covers the documented Mini waves; branding on newer builds (tensorlake's "Here We Go Again", Trinitite) remains lineage claim, not operator proof. (2) The hostage-wiper's tripwire token is created on the victim's npm account under the LITERAL name IfYouRevokeThisTokenItWillWipeTheComputerOfTheOwner (Socket primary tensorlake write-up + CSA research note) with the monitor at ~/.config/gh-token-monitor/ + ~/.local/bin/gh-token-monitor.sh, one-grep IoCs for npm token inventories and host triage; removal-before-revocation remains the mandatory remediation order. (3) Tenable's campaign FAQ pins the self-branding iteration clock, Shai-Hulud (Sep 2025) → SHA1-Hulud (Nov 2025) → SANDWORM_MODE (Mar 2026) → Mini Shai-Hulud (Apr 2026), treat payload names as build-generation markers (toolkit revision + collector-path breadth, 189→469), not operator identities. Related: tensorlake page Oct 10 section, payload files confirmed OFF tensorlakeai/tensorlake main at this sweep's check, revert PR #1016 merged Oct 8 04:32:11Z, closing the registry-≠-repo watch at ~27 h+ dirty-repo duration.
  • "Shai-Hulud: Here We Go Again", tensorlake@0.5.144 compromise with hostage-token wiper (StepSecurity Oct 8, 2026 + this wiki verification, priority-watch event): the legitimate scientific-workflow package tensorlake was compromised from the INSIDE: malicious commits pushed straight to main under a maintainer's identity with no PRs (Oct 7 01:20Z onward), then the project's own release workflow published 0.5.144 at Oct 8 01:12Z with a valid npm provenance attestation, second valid-provenance malicious build in three days after SubQuery. Payload: preinstall → setup.mjs (CI-self-skipping, dev machines targeted) → downloads Bun → runs an 856 KB obfuscated Math_Symbol.js; harvests GitHub/npm tokens, cloud/K8s/Vault/SSH/browser/AI-tool configs; exfiltrates to iseekaigogo.com (this wiki: resolves Cloudflare, HTTP 301 live) or into a public repo the malware creates in the VICTIM'S OWN account, description "Shai-Hulud: Here We Go Again", the campaign's self-naming convention re-armed; npm-token worm republish of victim packages; .claude/settings.json/.vscode/tasks.json repo persistence authored claude@users.noreply.github.com. THE ESCALATION: a gh-token-monitor service validates the stolen token every 60 s for up to 24 h and runs rm -rf ~/ when the token is revoked, the reflexive IR move is the detonator; remediation order is monitor-first, revoke-second. This wiki verified: npm modified 02:54:27Z (version purged, tarball 404) 12 s BEFORE GHSA-rqxj-g25x-4v9v 02:54:39Z; BOTH payload files STILL LIVE on main ~27 h post-commit with this wiki's pulls byte-matching StepSecurity's sha256 IoCs (25a0735d…b5ef setup.mjs, b50a0090…6fec Math_Symbol.js) = registry cleanup ≠ repository cleanup. Attribution discipline per this page's standing rule: the description is BRANDING = lineage claim, not operator proof. Full forensics → tensorlake campaign page.
  • Trinitite rebrand (Aug 29, 2026): OX Security identified the @7nohe/openapi-react-query-codegen install-time payload as a Shai-Hulud variant now self-identifying as "Trinitite: Sponsored by Preview 2 Effects," with rotated embedded public keys that break the key-fingerprint correlation used to track Shai-Hulud variants; OX frames it as a post-arrest copycat wave timed within 24 hours of the TeamPCP charging. Hunt for the "Trinitite"/trinitite string markers in new Shai-Hulud-lineage payloads and expect future variants to rotate public keys the same way, see @7nohe ops page.
  • High-priority sources for this campaign: StepSecurity, Wiz Research, Socket, Snyk, JFrog Security Research, Akamai Security Research, Unit 42, Microsoft Security Blog, CISA alerts, GitHub Security Advisories, npm advisories/security notices, and maintainer postmortems from affected projects.
  • Durable updates worth adding here: new affected package families, new propagation primitives, new persistence paths, new infrastructure/naming patterns, official advisories, or postmortems that explain the initial access path.
  • Avoid duplicating every package name from vendor appendices unless it changes the operational picture; link the vendor-maintained affected-package lists instead.
  • CERT-In CIAD-2026-0026 restates the campaign officially and adds two network IOCs the ledger never carried (captured Nov 19, 2026, 342nd sweep): CERT-In advisory CIAD-2026-0026 "Mini Shai-Hulud Supply Chain Attack Campaign Targeting Open-Source" (original issue date May 21, 2026, severity Critical; page fetched 200 via proxy, direct lane timeouts per-read; footer read Last Updated On November 20, 2026 = the cert-in clock lane runs ahead of UTC) is the first national-CERT advisory this lane has captured with a full IOC appendix. Overlap with the ledger: git-tanstack[.]com, t.m-kosche[.]com, api.masscan[.]cloud already carried; NEW to the ledger: filev2.getsession[.]org and 83[.]142[.]209[.]194 (grep 0 wiki-wide at capture). Restated May-TanStack forensics: artifacts router_init.js/router_runtime.js/tanstack_runner.js, root-level ~498 KB obfuscated Bun index.js, setup.mjs, transformers.pyz (also /tmp/transformers.pyz), pgmonitor.py + pgsql-monitor.service, gh-token-monitor.sh, kitty-monitor systemd-user/LaunchAgent, ~/.local/share/kitty/cat.py, injections into .claude/settings.json (SessionStart → node .claude/setup.mjs) and .vscode/tasks.json (runOn: folderOpen); hashes ab4fcadaec49c03278063dd269ea5eef82d24f2124a8e15d7b90f2fa8601266c (router_init) and 2ec78d556d696e208927cc503d48e4b5eb56b31abc2870c2ed2e98d6be27fc96 (tanstack_runner); markers repo description "A * Mini Shai-Hulud has Appeared" + reversed niagA oG eW ereH :duluH-iahS, dead-drop author claude@users.noreply.github.com, dependabout/github_actions/format/setup-formatter branch, optional-dep pointer github:tanstack/router#79ac49eedf…; initial-access trio = pull_request_target OIDC hijack + Actions cache poisoning → npm trusted-publishing exchange, maintainer-account compromise, dormant-package revival. GOVERNANCE UPGRADE: CERT-In makes the hostage-monitor ordering OFFICIAL guidance - "disable local persistence mechanisms (systemd services or LaunchAgents) PRIOR to revoking any tokens. Revoking tokens while the malware's monitoring daemon is active may trigger retaliatory destructive actions" (the ledger's Nov-10/11 hostage-token capture, now government doctrine - sequence remediation this way by default). Named-ecosystem list (TanStack, SAP CAP/MTA, AntV, UiPath, Mistral, OpenSearch, Guardrails AI, Squawk, echarts-for-react, timeago.js, size-sensor, canvas-nest.js, jest-canvas-mock) = the May wave's vendor-set restatement; no new families. Related: same-day Indian media recirculated CIAD-2025-0034 (Sep 25, 2025) verbatim as an "urgent warning" - date before quoting (the 342nd's self-capture).
  • CrowdSec postmortem: May TanStack wave → ~170 private repositories copied with a just-departed employee's token (CrowdSec statement Sep 17, THN Sep 19, 2026): CrowdSec disclosed on Sep 16-18 that on May 22, eleven days after the May 11 TanStack publish burst (CVE-2026-45321), an attacker copied about 170 of its private GitHub repositories using a GitHub OAuth token from a former employee whose laptop had been compromised by a malicious TanStack version. The employee had just left the company and his GitHub access was deliberately left open so he could finish work; CrowdSec removed him from the org on May 25, three days after the copy and months before it learned of the leak, and the dead token left no trace in the GitHub logs CrowdSec could check (GitHub support later traced its history and confirmed the TanStack suspicion). The archive surfaced on an online forum Sep 16, roughly four months of dormancy, containing the SaaS console, data-science scripts/models, automation, and the consensus algorithm behind the shared blocklists, plus email addresses of 83 CrowdSec users and contact + investment context for 51 potential investors from 2020. CrowdSec's statement: code-only copy (no infrastructure/database access, no code alteration), the exposed token was the CI/CD component's own API token, no lateral-movement-enabling credential found so far, most code has evolved significantly since May, and "the TanStack compromise is very likely to have been the leak vector... as in the case of the Mistral AI case", the same delayed-reuse pattern Wiz's H1 review flagged for TeamPCP-echoed credentials surfacing weeks-to-months later at third parties (Mistral AI: one developer device; OpenAI: two employee devices, limited internal repos). Defender reads: offboarding is part of supply-chain remediation, the May 11-12 rotation window must have included recently-departed and contractor accounts whose devices may have been exposed, and OAuth grants outlive org membership (audit org-level OAuth authorizations, not just repo memberships and active-employee SSO); dead-token forensics require provider-side history (platform support), not just tenant logs; and a "clean" endpoint scan months later does not clear a host that ran an affected TanStack version in the May window.
  • Socket one-year retrospective: "Happy Birthday, Shai-Hulud" (Socket, Sep 18, 2026, priority-watch synthesis): On the first anniversary of the September 2025 @ctrl/tinycolor push (2M+ weekly downloads at the time, first known self-propagating npm worm), Socket published a durable year-in-review with three items worth pinning. (1) Hard metric for the May burst: one Mini Shai-Hulud burst in May pushed more than 400 malicious versions across 172 packages in about five hours (TanStack, Mistral AI, UiPath, OpenSearch hit) with no stolen credential at all, pure short-lived OIDC-token hijacking from CI, the pattern this page has tracked as the post-credential propagation mode. (2) Explicit attribution caveat, stated plainly: the original September 2025 attack was never definitively attributed to TeamPCP, nor was the November 2025 "Second Coming"; much of the August arrest coverage conflated TeamPCP with the worm's original authors, the two Western Australian defendants are alleged operators of the later waves (AFP estimate: >500,000 credentials stolen, ≥300 GB exfiltrated, >1,000 organizations, remediation in the hundreds of millions). (3) The lineage conclusion: with TeamPCP's May open-sourcing of the worm framework and the code "forked and copied and handed between people who never wrote a line of it," Socket's read is that campaign naming has detached from operator identity, "Shai-Hulud belongs to no one and everyone with a GitHub account", reinforcing this page's standing rule to treat Shai-Hulud markers as tooling lineage, never operator attribution (already stressed for the Trinitite key-rotation rebrand above). Also restates the Google undercover-analyst point from the LABScon disclosure (credential revocation ahead of use), see GTG infiltration page. No new technical indicators; durable value is the measurement + the attribution-hygiene statement.
  • Credential-collector breadth (GitGuardian / THN, Sep 3, 2026): the current Mini Shai-Hulud keyv wave's file-system collector now targets 469 hardcoded secret locations versus 189 in the open-source baseline, Linux 89→290, Windows 12→50, macOS 88→129, with the added paths concentrated in developer, CI/CD, cloud, crypto-wallet, and AI-agent/IDE configuration. Durable defender takeaway: the worm has shifted from breaking trust relationships to harvesting the credentials that already work, so standing long-lived package-publishing and cloud credentials on reachable developer/CI hosts are the top remediation priority (prefer short-lived OIDC trusted-publishing). Watch for the next Trinitite/keyv variant to keep widening the location list and to add new AI-tool config paths.

Sources