harden-repo-for-coding-agents
Pass
Audited by Gen Agent Trust Hub on Jul 26, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill documentation in
references/playbooks/instruction-surface.mdcontains examples of common prompt injection patterns, such as "ignore previous instructions," which are used for educational purposes to guide users on auditing their repositories for security vulnerabilities. These strings are descriptive and do not present a risk of hijacking the agent. - [PROMPT_INJECTION]: The skill processes repository content to provide assessments, which constitutes an untrusted ingestion surface. The analysis identifies ingestion points (reading project files in
SKILL.mdsteps 2 and 4), notes that while playbooks recommend boundary markers for users the skill does not explicitly use them during its own ingestion, inventories capabilities (file writing viascaffoldanddiagnose), and highlights a self-audit sanitization step (SKILL.mdstep 5) where the agent re-reads diffs to verify compliance with safety guidelines. - [SAFE]: The skill demonstrates a strong security posture by advocating for deterministic enforcement mechanisms (e.g., PreToolUse hooks and CI gates) over purely instructional prose, and provides comprehensive guidance on secret management, egress control, and containerized sandboxing.
Audit Metadata