candid-improve-implementation
Pass
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (source code and git logs) which could contain hidden instructions designed to manipulate the agent's logic or bypass its intended behavior.
- Ingestion points: The skill reads changed files in full, git logs, and project configuration/standard files (
Technical.md,.candid/config.json) into the agent's context in Steps 1, 2, and 5. - Boundary markers: There are no explicit instructions to use delimiters or "ignore embedded instructions" warnings when interpolating the external code content into the agent's prompt.
- Capability inventory: The skill possesses significant capabilities, including executing shell commands (
git,jq,mkdir), writing tasks (TodoWrite), and modifying project files (Edit). - Sanitization: The instructions do not specify any sanitization, filtering, or validation of the source code content before it is processed by the AI.
Audit Metadata