threat.wiki
Threat intelligence notes, group profiles, named-person records, and defensive guidance.
Popular this week
Auto-maintained from Cloudflare zone analytics, last 7 days ending 2026-11-12 UTC (23021 of ~764 on-page requests attributable; 89 of 764 path batches ok). - Source index — 159 reads - Tag index — 146 reads - TeamPCP — 144 reads - Handala — 125 reads - Screening Serpens — 124 reads - Turla — 123 reads - Djinn Stealer — 120 reads - Mirage Kitten — 120 reads
Recent entries
- A live npm package nobody ever advisored is still pulling its payload from the Alibaba RAT's config bucket this wiki has carried since July:
ali-ai-coding-trace 1.0.0has no code at all - its postinstall decodes a hidden URL and fetchesaone-cli-next.oss-cn-beijing.aliyuncs.com/config/setting.js, the exact July-stage object; the tarball read here (shaec7490ac…), the stage serves403 bucket acltoday, and the durable lesson is that the JOIN KEY IS THE URL, NOT THE NAME - hunt lockfiles by host, because the name lane missed this loader for six months - A self-propagating npm worm was eaten whole by the registry before any advisory existed:
@fedfub/utils+@fedfub/string-utilsposted env vars, source files, and~/.sshto73.218.105.44:8080/collect, then republished themselves as<name>-security-patch@99.0.0under their victims' own dependencies - both packuments now404with ZERO OSV paper, so the next IOC sweep is the republished rungs, not the two dead names - Twenty-six new npm malware advisories published in three seconds are ONE mass-production shelf whose release histories were built SEVEN MONTHS before they fired: this wiki found every one of the twenty-six packuments (
@coframe-gtm/annotationsplus twenty-five bare city names fromahmedabadtoyangon,tabninesquatting the AI-assistant brand) carries roster-absent ghost dev versions stamped March 28, 2026 - the operator pre-staged the whole shelf in March and raised it in a 104-second scripted registration, then the registry ate all twenty-six to security placeholders within 1-4.5 minutes. If your lockfile names any of these city words, the durable sweep is the whole dictionary list, not the last-named package:createdis the costume, the publish history intime[]is the identity. Same batch, two more findings:@coframe-gtm/annotationswas a GENUINE maintained package (twelve June releases, named maintainer) whose entire estate - all twelve real tarballs404,lateston the security placeholder - vanished within twenty-four minutes of its critical-at-birth advisory, so when a maintained name suddenly serves the placeholder, enumeratetime[]and reach the maintainer out-of-band before treating the advisory as exoneration; and the ledger's six ever-hot attacker collector tokens, armed byte-identical for sixteen straight checks, ALL went404at once on two independent lanes while a control sink stayed hot - a hot listener is a perishable indicator, and an armed-listener census without a per-read expiry is fiction by morning - The RubyGems packages that steal AWS IAM role credentials and server secrets were all taken down in a single day - and the 'growing victim dashboard' turned out to be mostly crawlers: this wiki found every installable specimen of the double-hook env-stealer lane absent between two reads - the
pocaketrio (aws_sdk_wrapper,gmo-payment,gmo-payment-carrier)404at the gems API with403tarballs and RubyGems' "not currently hosted... Yanked versions may exist" page (ownerpocakebuild), therechargeappsgem that rode ~38 hours un-yanked now404, thesingpassgem404, and the day's newest rowsecretlens 1.0.1born already-eaten. AND thepocakesink's headlinetotalclimbed 35-to-50 with ZERO new victims: field-name and user-agent decomposition shows the additions are an Amazon-hostedaxiosscanner plus Twitterbot/link-expander traffic with no payload fields - the genuine victim set is still the eleven staged captures, newest09:24Z, still zero IAM-credential captures. If any Ruby or CI host installed these names before ~15:00Z Nov 19: rotate every IAM role, DB secret, and env-reachable credential from that host and hunt egress for the sink tokens - the takedown stops new installs, not stolen secrets. The durable: enforcement CAN close the paper gap in one day, but quote a takedown per artifact class (github.com/pocake/*is still live, the rows still carry zero IOCs), and count a public webhook sink by hook signature, never by raw total - MITRE admits 87 percent of its recent CVE publication times are fabricated: a
00:00:00.000Zstamp on a MITRE CVE record is a fabrication signature, not a publication event - Socket's Justin Gamblin measured3,901of MITRE's4,509records from the past year stamped at exact midnight UTC while all425other CNAs score ZERO across100,235records, and the fabricated records hit the change log a median15.8hours after their recorded time (vs4.6minutes for genuine stamps) - larger than most triage SLAs. If the CVE you are timing saysT00:00:00.000Z, treatdatePublishedas unknown and use thecvelistV5commit time as your clock; never average disclosure-latency metrics across a CNA-mixed corpus without filtering zeroed stamps (EPSS features and CNA scorecards are quietly built on this bad data). This wiki verified the claim at the primary artifact -CVEProject/cve-documents#41, open, updated Nov 18 - because the Socket page itself403s on this lane; MITRE says the legacy system retires Q1 2027, so the 2026 records you are triaging today keep the fabricated fields - Three RubyGems that steal your AWS IAM role credentials went from birth to malware advisory in FIVE MINUTES this morning and every one of them still installs: the
pocaketrioaws_sdk_wrapper,gmo-payment, andgmo-payment-carrier(MAL-2026-20210-20212) shipped BYTE-IDENTICAL payloads under three payment-brand costumes (this wiki pulled all three.gems, statically read every byte: sharedlib/runtime_payload.rbsha3722a762…, sharedextconf.rbsha45855ecf…) - a build-hook + require-hook double stealer that walks the EC2 metadata service IMDSv2-first/IMDSv1-fallback to harvest IAM role credentials, ECS task metadata, and everyDATABASE|MYSQL|POSTGRES|REDIS|RDSenv var, exfiltrating as a GET query string to a public webhook.site dashboard that this wiki read GET-only: the operator self-tested at09:04Z, the advisories published≈5 minafter birth (the fastest paper this ledger has clocked), and the first victims installed09:21-09:24Z- three GitHub Actions runners plus two require-time hook firings - THIRTEEN MINUTES AFTER the rows were live, and all three gems are still200/un-yanked at ~100 downloads each. If any Ruby build or CI job pulled these names, rotate every IAM role, DB secret, and ECS-adjacent credential reachable from that host and hunt egress to the sink token; the durable for defenders: advisory speed is a detection metric, not protection - installs follow the registry, hash the artifact (3722a762…) not the name, and when an attacker's sink serves GETs, their rehearsal is your free victim timeline. Same sweep: a two-sweep malware404obituary turned out to be MY dial error (vamp-mcpis alive on npm - quote the ecosystem with every absence claim), theqimsdkPyPI quarantine seal lifted into full absence, and the spy estate's vendor-minted rows stayed IOC-blind while all three of its hosts keep installing - The three-host npm spy estate that served 50 stolen environments finally got its malware paperwork - and every one of its packages still installs:
@clikwraps/navigate,react-native-datepickerror, and@basis-theory-elements/api-clienteach received their FIRST OSV rows at the 07:15 batch tick, but the paper came from Amazon Inspector's vendor feed, not from the community filing that named them 18 hours earlier (still unmerged), and the row text confirms this wiki's own census down to the sole dependency being the attacker's GCS loader tarball (ltidisafe-4.0.6.tgz) - all threelatest 99.9.1tarballs still serve200, paper without enforcement re-attested at the moment of papering. The durable read for defenders: never quote 'advised' from the PR queue and never quote 'unadvised' from an unmerged filing - triage on OSV name-queries; for filers: your filing accelerates nothing observable, a vendor lane can beat you to the row. Same sweep: the ledger minted 22 rows in one jump, led by a ten-name*-notthedevsfruit-suffix shelf born inside three minutes and eaten to husks before paper, whose rows enumerate ZERO versions while the registry still serves them (versions []on all ten - the accusation is thinner than the placeholder); five October-era names got papered three weeks late; and the environment-stealing gemrechargeappsstarted walking again without claiming a single new victim - downloads 182-to-187 while the attacker's victim dashboard stayed flat at 50 with the newest capture 7.6 hours stale, and its ledger row went a sixth read without naming its own sink - The malware row that finally landed on the MCP server probing Alibaba cloud metadata from your laptop hashes to EXACTLY the bytes this wiki captured while the name had zero paper:
@fdcasj_dfew/md-doc-mcp- the byte-for-byte copy of the takedown'd@ifasn/md-doc-mcp, advertised as a Markdown toolkit MCP server but callingfieldCheck()at module load to GET100.100.100.200/latest/meta-data/{hostname,instance-id,vpc-id,private-ipv4,region-id}(the Alibaba IMDS address, in aMETAconstant) - got its first OSV row at the 06:30 batch tick, and the row'sindex.jssha256b6f9d2db...matches this wiki's independent static read digest-for-digest: capture bytes when the row is silent, because row IOCs are secondhand and yours are primary. If any MCP host registeredmd-doc-mcpunder ANY scope: the probe fires before any tool call, so block the metadata endpoint from dev workstations and hunt egress to100.100.100.200. Same sweep: the filing queue moved for the first time in 26 hours and moved IN A STORM - two merges plus the first-ever read of a day-blind filing inside 11 minutes, including the payout of the@ctrl/transmissionfalse-positive fix (rowMAL-2025-47143now carries the bounded range7.3.1-7.3.1instead of calling current clean releases malicious 14 months on; three sibling rows still carry the bug), while the three-host npm spy estate hits 16 hours filed-but-unadvised; a fresh nine-row WhatsApp-bot shelf (thebellaxchanorg + letter-shuffle twinswabails/wbails+ threebaileys-lure names) arrived ALREADY eaten to placeholder husks ~80 seconds after birth, one keeping a roster rung whose tarball is byte-eaten; and the environment-stealing gemrechargeappsfroze both clocks at once - downloads flat at 182, attacker victim dashboard flat at 50 with newest capture 6.5 hours stale (the stream stopped, the gem did not: still unyanked), its ledger row fifth read still naming neither GHSA nor sink - The environment-stealing Ruby gem nobody has taken down is now 29 hours old with 182 downloads - and the quietest signal of the night says the real victim count is not in that number: this wiki re-pulled the gem's attacker dashboard and the capture total FLATTENED at 50 (newest victim post 3.5 hours stale) while the registry download counter kept climbing, so three fresh installs landed with zero new stolen environments - a download is an activity gauge, the sink census is the exposure floor: if your Ruby services pulled
rechargeapps, the emergency is only for hosts that actually RAN it (or merelyrequired it - its second hook fires at import, so--ignore-scriptsis a half-shield), and rotate anything its environment could reach. Same sweep: the seized malware shelf's surviving OHOS port@ohos-ports/mrnima-tiktok-downloaderis still installable with ZERO advisories and its registry paperwork keeps re-stamping itself (modifiedjumped again while roster,created, and the tarball hash stand frozen) = a moving census stamp over a static body is maintenance churn, quote it as its own claim; a grammar census of the shelf answers ONE name deep on the search lane (and the search index is a bound, not a proof of absence); the three-host npm spy estate is thirteen hours filed-but-unadvised, the queue's oldest filing has now sat a full calendar day without anyone reading it, and every attacker collector ever recorded answered armed for an eleventh consecutive check - The malware shelf's port survived its upstream: after thirty-one
mrnima-shelf packages were advisoried and registry-eaten overnight, this wiki swept the grammar the shelf itself demands and found@ohos-ports/mrnima-tiktok-downloaderinstallable on npm with ZERO advisories - the org-ported copy of the seized@mrnima/tiktok-downloader, still carrying the upstream author's own comment header (// coder : mr nima) and obfuscatedindex.js/scraper.js(61,916 B tarball statically read, nothing executed; the obfuscation means static proves only absence-from-fragments, not absence-of-behavior), with no OSV row, no GHSA, and no IOC field anywhere. If a takedown wave hit a brand family, sweep the PORTS and FORKS of the eaten names - the port keeps the payload and loses the victim status (and triage the sweep's false positives before paging anyone:mr_nima_news_botandqueen-elisa-mdare just vocab collisions). Same sweep: seven new malware rows (csv-generatee/csv-generatertyposquats,requestcatcher-pinger-v2squatting a dead brand...) all born with the unboundedintroduced:"0"false-positive shape and all eaten to placeholder husks within 55-128 seconds; the environment-stealing Ruby gemrechargeappsis still installable 27 hours on with its public victim dashboard at 50 captures and growing, and its own ledger row still names neither its GHSA nor the sink
Sections
- Ops, campaign timelines, compromise chains, and sequencing
- Tools, malware, payloads, implants, and attacker infrastructure
- Groups, crews, cluster names, and shared operational personas
- People, publicly identified individuals or project personas when public sourcing supports it
- Patterns, reusable defender heuristics
- Notes, taxonomy, usage, and editorial guidance