Skip to main content
Indicators are literal strings, such as file names, hosts and IP addresses, that do not appear in legitimate code. Behavior rules cover the techniques campaigns share, so new variants are caught even when the indicators change.

Tracked campaigns

Recurring techniques

The same handful of techniques shows up across campaigns. Each maps to a rule:
  • Code that runs without being imported. npm install scripts, Python .pth files and sitecustomize.py. See install-script, startup-code-execution and startup-hook.
  • Hiding in files CI already runs. Test setup, Makefiles and package scripts in a pull request, so the workflow itself looks unchanged. See CI file rules.
  • Stealing the workflow token. Reading the token actions/checkout leaves in git config, or requesting OIDC tokens to reach cloud accounts. See checkout-token-extraction and oidc-token-request.
  • Bringing a second runtime. Downloading Bun to run the payload outside Node.js tooling. See bun-dropper.
  • Harvesting with real tools. Running TruffleHog or Gitleaks on the machine or the repository history. See secret-scanner-download and history-secret-harvest.
  • Hiding secrets from log masking. Encoding secrets before printing them. See log-masking-bypass.
  • Exfiltrating to disposable infrastructure. Webhook services, paste sites, raw IP addresses and hostnames that embed an IP. See exfil-endpoint, bare-ip-request and secret-exfiltration.
  • Using the victim’s AI tools. Running coding agents with permission checks disabled, or planting autostart hooks for them. See ai-agent-abuse and agent-persistence.
  • Hiding the evidence. Reverting or force-pushing over the malicious workflow after it runs. See Scan recent runs.
  • Borrowing trust. Look-alike action owners, unclaimed owner names, and forks of popular actions. See the action rules in Workflow rules.

Sources