review
Warn
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from external, untrusted sources such as pull request titles, bodies, and linked specification documents. \n
- Ingestion points: Pull request metadata, linked URLs for specification documents, and issue/ticket content referenced within commit messages or descriptions (found in SKILL.md and pull-requests.md).\n
- Boundary markers: There are no instructions or requirements for the agent to use delimiters or system-level guardrails to ignore potentially malicious instructions embedded within the ingested external data.\n
- Capability inventory: The skill has significant capabilities, including shell execution via git, gh, and glab, as well as broad file system read access.\n
- Sanitization: The skill instructions lack any requirement for sanitizing or validating the content retrieved from external sources before it is processed by the agent.\n- [COMMAND_EXECUTION]: The skill provides shell command templates that incorporate user-supplied input directly into shell executions.\n
- Evidence: In github.md and gitlab.md, several command templates use placeholders like , , and directly in shell strings (e.g.,
gh pr view <pull-request-url>).\n - Risk: These patterns encourage a vulnerable execution model where user-controlled strings are interpolated into shell commands. If the agent does not implement its own robust escaping or validation, an attacker could provide a malicious URL or ID to execute arbitrary shell commands.
Audit Metadata