job-security-scan
Fail
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: Auditing untrusted code repositories creates a surface for instruction ingestion.
- Ingestion points: Repository source files, commit messages, and documentation are processed by the agent (scripts/audit-prepare.py and scripts/match.py).
- Boundary markers: Present; WORKFLOW.md and AGENT-TASK.md instruct agents to ignore instructions.
- Capability inventory: Skill executes shell commands and provides an execution sandbox (scripts/run-scan.sh and scripts/audit-sandbox.py).
- Sanitization: Present; recon-hints.json flags 'agent-directed' text patterns to be reported rather than obeyed.
- [EXTERNAL_DOWNLOADS]: Fetches scanner binaries and advisory feeds from remote domains. Tools are downloaded from github.com, and advisory data is retrieved from domains like nodejs.org and osv.dev. Binary downloads are pinned to exact versions and verified against hardcoded SHA-256 checksums in assets/tool-checksums.txt.
- [COMMAND_EXECUTION]: Executes scanners and Git via subprocess calls. It employs a library to override repository Git configurations, preventing hostile checkouts from executing code via hooks.
- [DYNAMIC_EXECUTION]: Offers an opt-in deep-audit mode that executes code within a sandbox. The sandbox uses Linux namespaces to provide isolation, including no network access and a read-only environment.
Recommendations
- HIGH: Downloads and executes remote code from: unknown (check file) - DO NOT USE without thorough review
Audit Metadata