continuous-improvement
Pass
Audited by Gen Agent Trust Hub on Apr 24, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface. It reads content from project-specific files and incorporates them into the prompts of spawned subagents without adequate sanitization.
- Ingestion points: In
SKILL.md, the orchestrator reads.claude/rules/continuous-improvement.mdand passes it to subagents. Additionally, subagents ingest project source code and runtime logs as part of the testing cycle described inreferences/testing-methodology.md. - Boundary markers: The instructions use markdown blocks but lack explicit "ignore embedded instructions" warnings or robust delimiters to prevent the agent from obeying commands found within the ingested project data.
- Capability inventory: The agents can spawn further subagents (
Agenttool), interact with GitHub repositories (ghCLI), and execute the project's own code (cargo run). - Sanitization: No validation or sanitization of external file content is performed before interpolation into agent prompts.
Audit Metadata