Skip to content

TraderTraitor / Jade Sleet

Summary

SentinelOne Labs published "Don't Call Us, We'll Call Your APIs: TraderTraitor Backdoors Resurface on Victim With No Crypto Ties" on September 18, 2026 (published alongside the authors' LABScon 2026 talk; The Hacker News picked it up September 21). It documents a second victim of the exact macOS backdoor pair used in the April 2026 KelpDAO/LayerZero $292M theft: an India-based IT services provider, unaffiliated with cryptocurrency, infected via the same DPRK fake-job-interview pipeline. The attribution sentence is now explicit: TraderTraitor aka UNC4899, PUKCHONG, Jade Sleet (the financially motivated Lazarus subgroup that hit Bybit via the Safe{Wallet} supply-chain compromise in early 2025 per GitHub's 2023 framing quote).

Three things make this durable intel rather than one more DPRK recruiting-scam writeup:

  1. FLATROOF = macOS.Gaslight. SentinelOne states the June 2026 Gaslight analysis (our macOS.Gaslight page) IS the FLATROOF backdoor under LayerZero/Mandiant's name, the first public alias join between the SentinelLABS tool family and the incident-named tooling.
  2. ROOFDECK's C2 bootstrap is a decentralized dead-drop: the implant needs a config file containing a Nostr public key of an attacker profile; on first run it pulls live relays from api.nostr[.]watch/v1/online, merges them with a hardcoded list of legitimate Nostr relays, finds the operator's profile, and uses the profile's website field as the C2 URL. All operator commands are signed with an attacker private key and signature-verified by an embedded public key before execution, C2 authenticity without infrastructure ownership, and command injection impossible for a defender who seizes the relay or C2 host.
  3. The disclosure-response behavior: on April 20, one day after LayerZero's public announcement, the attackers deployed a rebuilt ROOFDECK (stripped symbols/debug, new path, new pinned cert chain) that then deleted the old FLATROOF and ROOFDECK binaries. Vendor publication is an adversary trigger; assume your IR notes become attacker OPSEC input the moment they go public.

Tags

Why this matters

  • The developer endpoint is the access product. The victim was chosen the same way as LayerZero's "Developer1": an Apple Silicon MacBook whose owner ran Terraform/Ansible against AWS, OVH, and OpenStack, holding cloud credentials + source-control access. SentinelOne's formulation: individuals whose social media advertises infrastructure/DevOps work are candidates; the value of the target is whatever their laptop can reach.
  • Infrastructure-as-code files are now first-class phishing attachments. A terraform init on a repo from your "interviewer" treats .terraform.lock.hcl custom providers as the source of truth and downloads+executes attacker modules. The lockfile is trusted-by-design tooling nobody code-reviews.
  • Not every intrusion pays, and that changes triage. The attacker's own activity suggests the India victim "yielded insufficient value to sustain the intrusion", beaconing went intermittent after a C2 migration and the binary went to Trash by June 17. Low-and-slow doesn't mean abandoned-by-accident; it can mean downgraded-by-accounting, and can be resumed.
  • Signed commands + decentralized C2 discovery is the hardened-resilience pattern to design detections against, because takedown-centric defenses (sinkhole the C2, seize the domain) don't revoke the trust model.

Reported behavior

Delivery: fake interviews with per-victim backdoored repos

  • Classic Contagious Interview social engineering: the attacker contacts job seekers at the company that ends up compromised (not random candidates), each targeted profile being DevOps or crypto/FinTech engineering.
  • Lure repos are themed as infrastructure-engineering projects of the company the actor is impersonating; pivoting from LayerZero's published gtn-candidate-repo, SentinelOne found further lures referencing the companies Northwind and Novacart (fabricated or real-but-impersonated, unresolved).
  • Each repo ships a weaponized .terraform.lock.hcl with a custom provider pointing at attacker domains (three identified; SentinelOne's IoC list carries them). Running terraform init fetches and executes the malicious provider module.
  • Developer-awareness wrinkle: in one repo the candidate removed a typosquatted provider from the lockfile and noted it, possibly reading the infection as a security-awareness test. Lures now live inside a "would you run this?" teaching moment, which cuts both ways.

FLATROOF (aka macOS.Gaslight), initial-collector implant

  • Dropped as SystemUpdate in ~/Library/com.apple.iTunesCloud/SystemUpdate; ARM64 Rust.
  • On start it removes com.apple.quarantine from ROOFDECK and sets its executable bit, Gatekeeper is bypassed by the first-stage implant, so the second implant runs with no signature check and no prompt.
  • Detailed capability set + the analyst-targeting prompt-injection block are on our macOS.Gaslight page.
  • Ships with a Python collection module exfiltrating over Telegram using a built-in bot token.

ROOFDECK, the follow-on implant

  • Dropped as iSync in ~/Library/com.apple.internal.ck/iSync.
  • Requires a pre-existing config containing at minimum a Nostr public key → designed to be deployed only after foothold/control exists (~/.config/.repl_history stores C2 URLs, keys, beaconing settings, the filename itself is tradecraft: a fake REPL history).
  • C2 bootstrap: pull live relays from api.nostr[.]watch/v1/online + hardcoded legitimate Nostr relay list → locate the operator's Nostr profile by pubkey → read its website field = C2 URL. C2 endpoint path /app_version, HTTPS with an embedded pinned certificate.
  • Every C2 command is signed server-side; the implant verifies against an embedded public key before executing. Handlers are separated per command in source; the implant re-implements common shell commands for directory/file operations (SentinelOne ties this habit to sophisticated DPRK tooling incl. Lazarus' LightlessCan).
  • Persistence: ~/Library/LaunchAgents/loginwindow.plist with a dynamic app-identifier and a hardcoded --type=renderer argument to make the process look like a browser helper.
  • OPSEC leak: the pinned cert was made with mkcert default metadata, exposing a Linux QEMU VM running as username ub, SentinelOne found multiple attacker domains reusing that certificate CN = a certificate-name pivot that survives domain rotation.

Timeline on the India victim (telemetry-derived)

  • Backdoors on disk since March 18, 2026 (delivery vector unproven); dormant until March 29, when beaconing began, first launch by Cursor, seconds after the ~/DevOps-Automation/cloudshield workspace was opened. An IDE opening a workspace is the detonator; EDR that ignores cursor as "development noise" misses the parent-child signal.
  • April 20, 2026 (day after LayerZero's public KelpDAO disclosure): rebuilt ROOFDECK deployed as loginwindow at ~/Library/com.apple.appleaccountd/loginwindow (SHA-1 5728b11d30586bbfc1d8bd12df1c722a06e767a2), functionally identical but stripped of symbols/debug info, then used to delete both older binaries.
  • Intermittent beaconing to new C2 grenight[.]com until June 1; binary moved to Trash June 17, SentinelOne assesses the intrusion yielded insufficient value to sustain.

Indicators and hunt pivots

  • Paths: ~/Library/com.apple.iTunesCloud/SystemUpdate, ~/Library/com.apple.internal.ck/iSync, ~/Library/com.apple.appleaccountd/loginwindow, ~/Library/LaunchAgents/loginwindow.plist, config ~/.config/.repl_history.
  • Process tells: implant launched by Cursor/editor processes seconds after workspace open; LaunchAgent plists named loginwindow.plist with --type=renderer from outside a browser bundle; a binary stripping com.apple.quarantine off another file.
  • Network: HTTPS to /app_version with a pinned non-CA cert; DNS to api.nostr[.]watch (rare on corporate laptops unless the user is actually on Nostr); grenight[.]com; the three malicious Terraform provider domains in SentinelOne's IoC list (canonical page).
  • Certificate pivot: TLS certs issued by mkcert with default metadata / CN exposing username ub on a QEMU VM, hunt your TLS-inspection logs for that issuer string.
  • Hash: ROOFDECK rebuild SHA-1 5728b11d30586bbfc1d8bd12df1c722a06e767a2.
  • Supply-chain audit: any repo received out-of-band (interview projects, "take-home" repos) with a .terraform.lock.hcl naming a provider source that is not registry.terraform.io, treat terraform init on such a repo as executing attacker code.

Defender heuristics

  • Gate terraform init (and OpenTofu/Ansible Galaxy fetches) in CI and on managed endpoints to approved registry mirrors/provider domains; pin provider checksums and verify against the public registry.
  • Treat IDE-spawned first-execution of a previously-dormant binary as a Tier-1 detection, not development noise (parent = editor, child = unsigned binary in ~/Library/com.apple.*).
  • Alert on quarantine-attribute removal (xattr -d com.apple.quarantine) by non-installer binaries.
  • On any confirmed DPRK macOS implant: treat every cloud credential the laptop could reach as exposed (this actor's LayerZero objective was API keys → AWS/GCP privilege escalation), rotate at the identity layer, and audit for the follow-on implant rebuild, the attacker may already have replaced what your IR report described.
  • Search for attacker infrastructure by certificate identity (mkcert defaults, ub QEMU CN) rather than domains, this actor rotates domains within weeks.

Attribution notes

SentinelOne names the cluster TraderTraitor (aka UNC4899, PUKCHONG, Jade Sleet), the financially motivated DPRK Lazarus subgroup, and states cooperation with Google and Mandiant on the technical details. The June Gaslight assessment ("DPRK-aligned macOS activity cluster, high confidence") and this campaign-level naming are now the same lineage. SentinelOne's own caveat stands on the India victim: delivery mechanism unproven (implants on disk before observation began).

Sources