github-actions-script-injection
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill provides functional shell payloads for exfiltrating environment variables and system files (e.g.,
/proc/self/environ) to external URLs. These are presented as proof-of-concept examples for testing workflow security. - [INDIRECT_PROMPT_INJECTION]: The skill focuses on identifying attack surfaces in GitHub Actions workflow files where untrusted data is ingested. Specifically, it highlights the lack of boundary markers when context variables like
github.head_refare interpolated directly into shell steps. The methodology guides the agent to usefindandgrepto locate these unsanitized ingestion points. - [REMOTE_CODE_EXECUTION]: Payloads in the documentation demonstrate remote code execution patterns, such as piping
curloutput into abashshell to execute remote scripts. - [COMMAND_EXECUTION]: The audit methodology requires the agent to execute local shell commands (
find,grep) to analyze the repository's CI/CD configuration files for security weaknesses.
Audit Metadata