co-check
Fail
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The instruction to 'check that the cited... commands still succeed' for rows in SPEC.md requires the agent to execute arbitrary shell commands parsed directly from a markdown file.
- [REMOTE_CODE_EXECUTION]: Since SPEC.md is a project file that may be provided by untrusted contributors or fetched from remote repositories, executing the commands contained within it constitutes a remote code execution vector.
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing content from SPEC.md and using it to drive execution without sanitization. * Ingestion points: The skill reads SPEC.md and CHANGE.md files to identify tasks and associated commands. * Boundary markers: There are no markers or instructions to treat the file content as untrusted data. * Capability inventory: The agent uses command execution capabilities to verify the status of tasks cited in the specification. * Sanitization: No sanitization or verification of the command strings is performed before execution.
Recommendations
- AI detected serious security threats
Audit Metadata