impl-standards

Warn

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: MEDIUMPROMPT_INJECTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill includes 'Self-Evolving' instructions that direct the agent to modify the SKILL.md file based on its own runtime performance or observed issues. This pattern allows for autonomous modification of the agent's core instructions, which can lead to instructional drift or unauthorized behavioral changes.
  • Evidence: 'This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer.' in SKILL.md.
  • Evidence: 'Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.' in SKILL.md.
  • [DYNAMIC_EXECUTION]: The documentation provides examples of executing scripts where the file path is determined at runtime using environment helpers. Executing code from computed paths can be risky if the path resolution logic or the target directory is compromised.
  • Evidence: ROOT="$(cc-plugin-root itp)" and uv run --script "$ROOT/skills/code-hardcode-audit/scripts/audit_hardcodes.py" -- src/ in references/constants-management.md.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external source code files using Read, Grep, and Glob tools. It lacks clear boundary markers to distinguish between the content being analyzed and the agent's instructions, creating a surface where malicious instructions embedded in code (e.g., in comments) could influence the agent's behavior.
  • Ingestion points: The skill uses Read, Grep, and Glob to analyze project files in the src/ directory.
  • Capability inventory: The agent is authorized to modify its own instructions (SKILL.md) and is shown how to execute shell commands for auditing.
  • Boundary markers: No explicit delimiters or instructions are provided to the agent to treat analyzed code strictly as data and ignore any embedded natural language instructions.
  • Sanitization: No sanitization or validation of the content read from files is performed before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 8, 2026, 02:14 AM