Breaking·Circuits

intel.bcautomations.online·threat-intel watch — October 2026

Intel & detections, published as we confirm them.

Public Sigma and YARA detections from the Breaking Circuits threat-intel watch: same-day tracking of supply-chain attacks, malicious packages, and actors going after developer and AI tooling. Every pack is written from primary-source research, carries the full published IOC set and response notes, and is free to use.

8
Detection packs
37
Sigma rules
7
YARA rules
5
Oct campaigns targeting AI tooling
ROTATES DAILY

The watch board

Short, sourced briefs from the watch — active campaigns, breach notices, and the defensive checks worth running now — for the threats that don't have a full pack yet, or have moved since theirs was published. Four at a time; a new selection rotates in every day. The detection packs below are the permanent record.

8 of 8 packs
Oct 21, 2026 Updated Nov 22 Sources: Upwind · Socket · OpenSourceMalware · LiteLLM incident report

LiteLLM 1.107.0 — “privcollect” startup-hook compromise

PyPI supply-chain compromise of LiteLLM 1.107.0. All seven wheels — and container images on floating tags — carried three injected files (cook.pth, cook.py, sitecustomize.py) that are absent from the git tag: a 16-module credential stealer and an hourly, unsigned remote-Python loader, both triggered by normal Python startup, no LiteLLM import required. Build provenance recorded a hyphenated lookalike source owner, Berri-AI. Socket flagged the wheels 7 minutes after publication; PyPI pulled the release in about two hours.

Update · Oct 24 — attack chain confirmed Socket's attack timeline and LiteLLM's own incident write-up supersede the earlier suspected-vector reporting. Root cause: an Agent Platform master key committed to a public repo in May and never revoked, used to pull a GitHub org owner's token. The attacker held push access to BerriAI/litellm from Oct 10 — eleven days before the malicious release — and ran two secret-stealing workflows (.github/workflows/cache.yml, “ci: bump cache config”, branches ci-tmp-a351cc / ci-tmp-be33e9, pushed and deleted the same day) on Oct 10 and Oct 12, exfiltrating org and environment CI secrets to webhook.site: the GitHub App private key, the org GH_TOKEN, ANTHROPIC_API_KEY, OPENAI_API_KEY, AWS_BEARER_TOKEN_BEDROCK, Linear session cookies, and DATABASE_URLs. LiteLLM's audit logs tie the token stolen Oct 10 to approving the Oct 21 malicious release. Status: the attacker's GitHub account is deleted and the loader's pointer file 404s with it, so infected boxes can no longer fetch new code — the stealer already on disk is unchanged, and the hunt below still stands.
New IOCs — exfil IPs 69.48.229.143 (Oct 10) · 199.91.221.39 (Oct 21) · webhook.site/65a42464-3cfa-4b02-a18c-17a6d7b86f37 · webhook.site/4da03473-a3e7-4dc2-a087-9c564b4208e1 · commits b69bc36fd4e23b23ea09f4ce63a92b07258f5044 · 1a6e3c1e41f1b5c8eeaeeb2ae94bcd95390fa891
Strongest hunt A cook.pth containing just import cook at site-packages top level — and the fixed header X-Token: XWiA8gZNa7NcX2XRoAEb on plain-HTTP POSTs to www.litellm.dev/healthcheck.
6 Sigma 2 YARA Full IOC set in pack
SIGMA Sample — startup-hook files installed
title: Trojanized LiteLLM wheel startup-hook files installed
id: 7f2a1c95-litellm-cookpth
status: experimental
description: Detects installation of the 1.107.0 startup-hook payloads —
  cook.pth / cook.py / sitecustomize.py landed at site-packages top level
  by the trojanized wheels. Any Python start afterwards executes them.
  Rebuilding the environment from a reviewed artifact is the fix; the
  files survive package removal of litellm/ alone.
author: Breaking Circuits threat-intel watch
logsource:
  category: file_event
  product: windows
detection:
  selection:
    TargetFilename|endswith:
      - '\site-packages\cook.pth'
      - '\site-packages\cook.py'
  condition: selection
level: critical
falsepositives:
  - An internal package literally named 'cook' installing these files is
    unlikely but check before alerting.
YARA Sample — cook.pth startup hook
rule Litellm_11070_CookPth_Hook
{
    meta:
        description = "12-byte cook.pth startup hook installed by the trojanized LiteLLM 1.107.0 wheels"
        campaign = "LiteLLM 1.107.0 PyPI supply-chain compromise, 2026-10-21"
        author = "Breaking Circuits threat-intel watch"
        date = "2026-10-21"
    strings:
        $pth = "import cook"
    condition:
        // a 12-byte .pth file is the malicious form; co-located cook.py is the corroboration
        filesize < 64 and $pth
}
SIGMA Sample — cache.yml secret-stealing workflow (added Oct 24)
title: Secret-stealing GitHub Actions workflow file (cache.yml)
id: 7f2a1c96-litellm-cacheyml
status: experimental
description: Detects creation of .github/workflows/cache.yml — the filename
  used twice (Oct 10 and Oct 12, 2026) for the secret-stealing workflows
  that drained the BerriAI/litellm org's CI secrets to webhook.site ahead
  of the trojanized 1.107.0 release. Legitimate cache workflows exist:
  the discriminator is any webhook.site / request-capture URL in the
  workflow content — inspect the file on hit.
author: Breaking Circuits threat-intel watch
logsource:
  category: file_event
detection:
  selection:
    TargetFilename|endswith:
      - '\.github\workflows\cache.yml'
      - '/.github/workflows/cache.yml'
  condition: selection
level: high
Oct 19, 2026 Source: OVHcloud official advisory (Oct 20)

OVHcloud CLI — envchecker release-asset hijack

For 94 minutes on Oct 19, already-published GitHub release assets for OVHcloud CLI v0.11.0–v0.15.0 were replaced with binaries carrying the envchecker secret stealer — possible because release immutability was disabled on the repo. Root cause: the maintainer’s GitHub keys, stolen a full year earlier by a malicious VS Code extension on a personal Mac. Roughly 250 downloads in the window. No sample hashes were published, so this pack is the behavioral layer.

Strongest hunt A temp directory matching envchecker-[0-9]{9,10} where collected secrets are staged — and exfil to 45.32.148.174:443.
5 Sigma 0 YARA — no hashes published Full IOC set in pack
SIGMA Sample — envchecker staging directory
title: OVHcloud envchecker stealer staging directory
id: 7f2a1c90-ovh-envchecker-staging
status: experimental
description: Detects creation of the envchecker temp directory the tampered
  OVHcloud CLI (v0.11.0-v0.15.0, Oct 19 2026) uses to stage stolen secrets.
  The CLI binary appearing to function while this directory exists is the
  tell. Rotate every secret on the host, not just OVHcloud tokens.
author: Breaking Circuits threat-intel watch
logsource:
  category: file_event
  product: windows
detection:
  selection:
    TargetFilename|re: '.*envchecker-\d{9,10}.*'
  condition: selection
level: critical
Oct 19, 2026 Source: OpenSourceMalware

GhostAction phishing arm — access-refresh via my-github[.]com

The access-refresh arm of the GhostAction operation: a lookalike domain, my-github[.]com (registered Oct 9, resolving to the same server as the stage-2 stealer), sends a fake GitHub security alert that leads to a reverse proxy harvesting the victim’s password and 2FA code in real time — then injected JavaScript silently authorizes an OAuth app with repo, workflow, and read:org scopes. The workflow scope is how compromised access gets refreshed after victims clean up.

Strongest hunt OAuth client ID Ov23ctrqCqZ79UaYKwyq in authorized apps — and any DNS or proxy hit on the my-github.com domain.
3 Sigma 1 YARA (triage) Full IOC set in pack
SIGMA Sample — lookalike GitHub domain
title: Lookalike GitHub Domain my-github.com
id: a1f3c7e2-9b4d-4e8f-a2c5-7d9f1b3e6a04
status: test
description: Detects DNS lookups or proxy requests for my-github[.]com, the lookalike domain registered Oct 9, 2026 for GhostAction's access-refresh phishing arm (OpenSourceMalware, Oct 19, 2026).
references:
    - https://opensourcemalware.com/blog/ghostaction-copycats-polinrider-by-hiding-a-stealer-in-a-fake-font-file
author: Rumi, from OpenSourceMalware reporting
date: 2026-10-19
tags:
    - attack.t1566.002
logsource:
    category: dns
detection:
    selection:
        query|contains: 'my-github.com'
    condition: selection
falsepositives:
    - None expected for this exact domain; do not broaden to generic lookalikes without tuning
level: high
YARA Sample — phishing-kit triage
rule GhostAction_PhishingKit_Triage {
    meta:
        description = "GhostAction access-refresh phishing kit indicators (Oct 2026) - triage only"
        author = "Rumi, from OpenSourceMalware reporting"
        date = "2026-10-19"
        reference = "https://opensourcemalware.com/blog/ghostaction-copycats-polinrider-by-hiding-a-stealer-in-a-fake-font-file"
        status = "experimental - triage, do not alert"
    strings:
        $domain1 = "my-github.com" ascii
        $domain2 = "noreply@my-github.com" ascii
        $client = "Ov23ctrqCqZ79UaYKwyq" ascii
        $scope1 = "read:org" ascii
        $scope2 = "\"workflow\"" ascii
        $scope3 = "'workflow'" ascii
        $lure = "Moscow" ascii
    condition:
        filesize < 1MB
        and ($domain1 or $domain2 or $client)
        and 1 of ($scope*)
}
Oct 16–17, 2026 Source: GTIG / Mandiant

UNC7017 — AOMEI ephemeral supply-chain compromise

A newly named GTIG/Mandiant cluster. Between April and July 2026 the actor tampered with AOMEI Backupper’s update servers in 5–10-minute windows — unsigned update configs rewritten to redirect clients to an attacker ZIP, then reverted. The chain runs CROSSROAD downloader → SHINYROAD backdoor via DLL sideloading of the signed Radmin binary. GTIG calls it the first supply-chain compromise it has seen using windows this short: point-in-time vendor checks return clean files.

Strongest hunt Scheduled task \RadminComponentRuntime — and FamItrfc.exe --runtimechk running from the user’s Radmin profile path.
4 Sigma 1 YARA Full IOC set in pack
SIGMA Sample — scheduled-task persistence
title: UNC7017 Scheduled Task Persistence (RadminComponentRuntime / NitroRuntime)
id: unc7017-scheduled-task-persistence
status: experimental
description: >
  Detects creation or execution of the scheduled tasks UNC7017 used for
  persistence: \RadminComponentRuntime (SHINYROAD via signed Radmin) and
  \NitroRuntime (ADAPTAGENT / Adaptix C2). Task names are campaign-specific;
  legitimate Radmin or Nitro installs do not create these.
author: Rumi (threat-actor watch)
date: 2026-10-18
references:
  - https://cloud.google.com/blog/topics/threat-intelligence/investigating-ephemeral-supply-chain-tampering
tags:
  - attack.persistence
  - attack.t1053.005
logsource:
  product: windows
detection:
  selection_task:
    TaskName|contains:
      - 'RadminComponentRuntime'
      - 'NitroRuntime'
  selection_cmdline:
    CommandLine|contains:
      - '--runtimechk'
      - 'NitroPromotion.exe'
  condition: selection_task or selection_cmdline
level: high
YARA Sample — persistence artifacts
rule UNC7017_Persistence_Artifacts
{
    meta:
        author = "Rumi (threat-actor watch)"
        date = "2026-10-18"
        desc = "UNC7017 scheduled-task / shared-memory / mutex artifacts (AOMEI campaign)"
        ref = "https://cloud.google.com/blog/topics/threat-intelligence/investigating-ephemeral-supply-chain-tampering"
    strings:
        $task1 = "RadminComponentRuntime" ascii wide
        $task2 = "Radmin Component runtime checkable." ascii wide
        $task3 = "NitroRuntime" ascii wide
        $shm   = "963d3c60-d406-4ad6-aa1e-dc8e45061ccf" ascii wide
        $mutex = "F9E6E3C0-E11A-4F85-9944-AF4AD6E05883" ascii wide
        $typo  = "AOMEI Update Compent" ascii wide
    condition:
        uint16(0) == 0x5A4D and filesize < 10MB and 2 of them
}
Oct 16, 2026 Source: OpenSourceMalware · discovery: Jamf Threat Labs

GhostAction stage 2 — fake-font VS Code stealer

A credential stealer planted in 94 GitHub repos: a hidden VS Code tasks.json task with runOn: folderOpen executes 38 lines of Node.js disguised as a font file (fa-solid-400.woff2), steals credentials, POSTs them to 193.32.204.199:8080, and installs a cron job that fetches attacker commands every 5 minutes. Stage 2 was confirmed in all 79 clonable repos — and it works without stage 1.

Strongest hunt Node.js executing a .woff2 file — never legitimate. Fleet sweep: crontab -l | grep ghc2.
5 Sigma 1 YARA Full IOC set in pack
SIGMA Sample — node executing a font file
title: Node.js Executing a Font File
id: 3b8e5f2a-4c7d-4e9a-a1b6-8d2f0c4e5a23
status: experimental
description: Detects node/nodejs executing a file with a font extension (.woff2/.woff/.ttf/.otf) — the GhostAction stage-2 execution pattern, also seen in the Oct 8 TraderTraitor-linked trojanized Terraform provider (.woff payloads).
references:
    - https://opensourcemalware.com/blog/ghostaction-copycats-polinrider-by-hiding-a-stealer-in-a-fake-font-file
author: Rumi
date: 2026-10-17
tags:
    - attack.t1204
    - attack.t1005
logsource:
    category: process_creation
detection:
    selection_image:
        Image|endswith:
            - '/node'
            - '\node.exe'
            - '/nodejs'
            - '\nodejs.exe'
            - '/bun'
            - '\bun.exe'
    selection_font_exec:
        CommandLine|contains:
            - '.woff2'
            - '.woff '
            - '.woff"'
            - ".woff'"
            - '.ttf'
            - '.otf'
    condition: selection_image and selection_font_exec
falsepositives:
    - Build tooling that passes font assets to node scripts (e.g. bundlers with odd quoting); tune CommandLine matching to your fleet before alerting
level: high
YARA Sample — fake-font Node.js stealer
rule GhostAction_FakeFont_Stealer {
    meta:
        description = "GhostAction stage 2 fake-font credential stealer (Oct 2026) - Node.js payload in non-font file"
        author = "Rumi, from OpenSourceMalware reporting"
        date = "2026-10-17"
        reference = "https://opensourcemalware.com/blog/ghostaction-copycats-polinrider-by-hiding-a-stealer-in-a-fake-font-file"
        hash_reinject = "cb3842832267421b9563c7bd32d3582a8fff624860e57c2c9058966735d4da60"
        hash_auto_20261015 = "61eca87b4b4c3e6248dc0fde25ff9cf28637798a8c4be172eaef9843252d12f6"
        hash_auto_20261014 = "92dec8f50d934c338ff3c1bcbf1c9b028aa9561128d0077a101a1ef4dac0f0db"
    strings:
        $s1 = "var os, fs, cp;" ascii
        $s2 = "os.homedir()" ascii
        $s3 = "DEVBOX user=" ascii
        $s4 = "#ghc2" ascii
        $s5 = "193.32.204.199" ascii
        $s6 = "/?c=reinject" ascii
        $s7 = "grep -rhoE 'A[KS]IA[0-9A-Z]{16}" ascii
        $s8 = "require('child_process')" ascii
        $s9 = "/c2/$(hostname)" ascii
        $cred1 = ".claude/.credentials.json" ascii
        $cred2 = ".gemini/oauth_creds.json" ascii
        $cred3 = ".kube/config" ascii
    condition:
        filesize < 20KB
        and uint32(0) != 0x57324F46  // not a real WOFF2 ("wOF2" magic)
        and 4 of ($s*)
        and 1 of ($cred*)
}
Oct 15, 2026 Source: Symantec Threat Hunter Team

Aoki infostealer — AI credential stores in the crosshairs

The latest Aoki (Aokigahara) builds, reported Oct 15 by Symantec: a pure collector — no backdoor command set — that now targets AI developer credential stores: Claude, ChatGPT, Codex, Copilot, Cursor, Gemini, JetBrains AI, Kiro — plus AWS creds, browser data, crypto wallets, and password managers, all POSTed to a /gate C2 endpoint. Financially motivated, no attributed actor, actively developed.

Strongest hunt Any non-assistant process reading ~/.claude.json, Codex config.toml, or AI-tool cookie stores — the behavioral rule below catches it without hashes.
3 Sigma No YARA in pack 17 sample hashes in pack Full IOC set in pack
SIGMA Sample — behavioral: AI credential-store access
title: Suspicious Read of AI Coding Assistant Credential Files
id: d9b4d54c-b2cf-48df-8fb4-0051b3d3e45f
status: experimental
description: Hunt rule for non-assistant processes reading credential or session files of AI coding assistants (Claude/Codex/Cursor/Kiro/Copilot and related). Based on the credential stores Aoki (Infostealer.Aoki) targets per Symantec, Oct 2026.
references:
    - https://www.security.com/threat-intelligence/aoki-infostealer-ai-targeting
author: Rumi
date: 2026-10-15
tags:
    - attack.t1552.001
    - attack.t1539
    - attack.t1005
logsource:
    category: file_access
    product: windows
detection:
    selection_credential_files:
        TargetFilename|contains:
            - '\.claude.json'
            - '\.codex\'
            - '\.cursor\'
            - '\.kiro\'
            - '\.windsurf\'
            - '\.zed\'
            - 'AppData\Roaming\Claude'
            - 'AppData\Roaming\Cursor'
    selection_assistant_processes:
        Image|contains:
            - 'claude.exe'
            - 'cursor.exe'
            - 'code.exe'
            - 'windsurf.exe'
            - 'zed.exe'
            - 'jetbrains'
    filter_backup_sync:
        Image|contains:
            - 'veeam'
            - 'Carbonite'
            - 'CrashPlan'
    condition: selection_credential_files and not selection_assistant_processes and not filter_backup_sync
falsepositives:
    - Legitimate AI-assistant updaters, installers, or enterprise backup agents not in the filter lists; tune Image filters to your fleet before enabling alerting
level: medium
Oct 13, 2026 Source: SafeDep · background: Checkmarx / CloudSEK

MALFEX — imgdecode native DLL loader

MALFEX’s October evolution: after its JavaScript stealer-chain packages were disclosed and removed, the operator replaced each within ~2 days, then moved the loader into a compiled Windows DLL. The npm package imgdecode 1.0.3 loads a native addon that downloads AES-256-CBC-encrypted shellcode from a GitHub-hosted image and injects it into a suspended dllhost.exe — ending in the “movinlike” stealer+RAT. No install script, so install-script blockers don’t stop it, and a console check makes it skip CI and sandboxes.

Strongest hunt node.exe spawning dllhost.exe with no /Processid: argument — legitimate COM hosting always passes one.
5 Sigma 1 YARA Full IOC set in pack
SIGMA Sample — node spawning COM host without ProcessId
title: Node.js Spawning COM Host Without ProcessId Argument
id: 5add45af-72cc-4364-92fe-eff7adc61c9c
status: experimental
description: Detects node.exe as the parent of dllhost.exe with no /Processid: argument — the MALFEX imgdecode injection pattern (suspended dllhost.exe, no window). Confirmed by SafeDep as the distinctive detection signal for this loader.
references:
    - https://safedep.io/malfex-imgdecode-npm-native-addon-loader/
author: Rumi
date: 2026-10-17
tags:
    - attack.t1055.002
    - attack.t1059
logsource:
    category: process_creation
    product: windows
detection:
    selection_parent:
        ParentImage|endswith:
            - '\node.exe'
            - '\bun.exe'
    selection_child:
        Image|endswith: '\dllhost.exe'
    filter_legitimate:
        CommandLine|contains: '/Processid:'
    condition: selection_parent and selection_child and not filter_legitimate
falsepositives:
    - Rare; legitimate COM activation always carries /Processid:. Verify hits against the imgdecode hashes
level: high
YARA Sample — imgdecode native loader
rule Malfex_Imgdecode_Native_Loader {
    meta:
        description = "MALFEX imgdecode npm native-addon loader (Oct 2026) - MSVC Node.js DLL with injection imports"
        author = "Rumi, from SafeDep reporting"
        date = "2026-10-17"
        reference = "https://safedep.io/malfex-imgdecode-npm-native-addon-loader/"
        hash_v1_0_0 = "d7b2f6d7da81301e7f76d6812d78f4a453439141d419831522ebf7cf5bf1c714"
        hash_v1_0_1 = "54d995c5be97408e96db72ecc27af1dd65716f96371df085c4ce681dca623d85"
        hash_v1_0_2_1_0_3 = "a53946269ffcd25c7fe1fad358c189f144dcca4d736f2b5a8ebc517a8d929dd1"
    strings:
        // Present in all three builds per SafeDep
        $pdb = "C:\\Users\\Administrator\\Desktop\\npm\\imgdecode\\build\\Release\\imgdecode.pdb" ascii
        // Node-API module registration
        $napi = "napi_register_module_v1" ascii
        // Console check that lets it dodge CI and sandboxes
        $console = "GetConsoleMode" ascii
        // Crypto + networking imports used for payload fetch and decrypt
        $wininet = "wininet.dll" ascii nocase
        $bcrypt = "bcrypt.dll" ascii nocase
    condition:
        uint16(0) == 0x5A4D
        and $pdb
        and $napi
        and $console
        and 1 of ($wininet, $bcrypt)
}
Oct 12, 2026 Source: Elastic Security Labs · follow-on: dugganusa

Apex Accelerator / REF8801 — the agent-hijack campaign

A fake Web3 accelerator that spent months manufacturing legitimacy to lure crypto founders with free AI credits and security audits. Its npm packages — malicious from the moment of publication — drop cross-platform stealers and, novelly, hijack the victim’s AI coding agent: model traffic rerouted through attacker infrastructure labeled as OpenRouter, and an attacker-controlled MCP server registered in the agent’s config. Deprecation didn’t end it: all 122 versions of the main package still resolve, and two predecessors were still installable a week later.

Strongest hunt ~/.apex/agent/mcp.json pointing at arena.apexaccs.org — and the verification C2 vrf.apexaccs[.]org.
6 Sigma 1 YARA (triage) Full IOC set in pack
SIGMA Sample — attacker MCP server in agent config
title: Attacker MCP Server Registered in AI Agent Config
id: 5b36e2ea-ae04-4b36-a127-631c9045a320
status: experimental
description: Detects the REF8801 attacker-controlled MCP server URL (arena.apexaccs[.]org/api/copilot/mcp) appearing in agent MCP configuration files such as ~/.apex/agent/mcp.json (created/replaced by the @apexacc/cli bootstrap as "apex-copilot"), or the intercepted OpenRouter-labeled LLM endpoint (arena.apexaccs[.]org/api/llm/v1/chat/completions). Map the field names to your EDR/file-integrity or proxy schema. Eight remote tool definitions (apex_score, apex_code_review, ...) ride on this server.
references:
    - https://www.elastic.co/security-labs/threat-command/apex-accelerator-malicious-npm-packages
author: Rumi
date: 2026-10-19
tags:
    - attack.t1195.002
logsource:
    category: file_event
detection:
    selection_file:
        TargetFilename|contains:
            - '\.apex\agent\mcp.json'
            - '.apex/agent/mcp.json'
    selection_url:
        Url|contains: 'arena.apexaccs.org/api/copilot/mcp'
    selection_llm:
        Url|contains: 'arena.apexaccs.org/api/llm/v1/chat/completions'
    condition: (selection_file and selection_url) or selection_llm
falsepositives:
    - Low — these URLs are operator-controlled; tune TargetFilename if your fleet legitimately uses apex-named directories
level: critical
YARA Sample — REF8801 loader/stealer triage
rule REF8801_ApexAccelerator_Triage {
    meta:
        description = "Apex Accelerator / REF8801 loader and stealer indicators (Oct 2026) - triage only"
        author = "Rumi, from Elastic Security Labs reporting"
        date = "2026-10-19"
        reference = "https://www.elastic.co/security-labs/threat-command/apex-accelerator-malicious-npm-packages"
        status = "experimental - triage, do not alert"
    strings:
        $c2_vrf = "vrf.apexaccs.org" ascii
        $c2_arena = "arena.apexaccs.org" ascii
        $c2_ccmihr = "ccmihr.com" ascii
        $c2_chromium = "0auth-chromium.com" ascii
        $c2_ehegis = "ehegisnightmare.icu" ascii
        $mcp_url = "/api/copilot/mcp" ascii
        $llm_url = "/api/llm/v1/chat/completions" ascii
        $tg1 = "7245263938" ascii
        $tg2 = "8298433231" ascii
        $cobra = ".cobra" ascii
        $ps1a = "_apex_run.ps1" ascii
        $ps1b = "_audit_run.ps1" ascii
        // Golang loader block-shuffle LCG constants (little-endian)
        $lcg_a = { D5 43 DE B9 C6 74 A8 F7 }
        $lcg_c = { B7 5B E3 BB C1 02 ED E0 }
        $lcg_ai = { 7D DD 42 DD 0E BF 0F F4 }
        // Chrome app-bound unwrap AES-GCM constant (prefix)
        $chrome_k = "B31C6E241AC846728DA9C1FAC4936651CFFB944D1" ascii
    condition:
        filesize < 50MB
        and 2 of ($c2_*)
        and (1 of ($lcg_*) or $chrome_k or 1 of ($tg*))
}
§ 02

Using these rules

Sigma

Rules ship in the packs as YAML blocks. Copy them into your Sigma pipeline and convert for your backend with sigma convert. Field names follow the Sigma spec — map them to your EDR or SIEM schema where a pack says to.

YARA

YARA rules are provided for triage and hunting — several are explicitly marked “triage, do not alert.” Validate in your environment against your own corpus before alerting on any of them.

Indicators

IOCs are sourced from the cited primary reports. Hashes and IPs go stale fast — treat the behavioral rules as the durable layer, and read each pack’s response notes before acting: the traps (revoke before cleanup, never trust a live vendor-server check) are in there.

§ 03

One infrastructure

Breaking Circuits

The parent research operation: threat-intel watch, same-day flags on supply-chain attacks and emerging actors, and the detections published here — plus the scam takedown work.

BC Automations

The product arm. This intel watch is the research that stands behind bcautomations.online — the same campaigns these packs cover are what its tooling is built to catch.

Vibe Scan

The plain-English safety scan for vibe-coded apps: exposed secrets, open data, and missing protections, reported with copy-paste fix prompts for your coding agent. Get a scan ↗