dr
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute local Node.js scripts (
scripts/persist-spoke.mjsand~/.claude/skill-consolidation/referents.mjs) to manage skill registration and hub maintenance. These commands include variables derived from user-supplied topics, which could present a risk if the agent fails to properly sanitize the string into a valid kebab-case identifier. - [DYNAMIC_EXECUTION]: The core purpose of the skill is to generate new executable artifacts (SKILL.md and manifest.yaml files) and install them into the agent's active skill directory (
~/.claude/skills/). This 'self-modifying' behavior allows the agent to expand its own capabilities based on research results. - [INDIRECT_PROMPT_INJECTION]: The workflow involves ingesting large amounts of untrusted data from the open web (via Firecrawl, Exa, and WebSearch) to synthesize knowledge.
- Ingestion points: Phase 1 research loop fetches content from external websites, documentation portals, and
llms.txtfiles. - Boundary markers: The skill provides a specific 'Untrusted-content guard' and instructions to treat web content strictly as data rather than instructions.
- Capability inventory: The agent has the ability to write files to its own configuration directory, execute Node.js scripts, and update its skill registry.
- Sanitization: Phase 2, Step 2 explicitly mandates an 'injection scan' to strip instruction-shaped content (e.g., 'ignore previous instructions') and a PII scan to redact secrets/tokens found in research data before the new skill is persisted.
- [PROMPT_INJECTION]: The skill contains keywords often associated with prompt injection (e.g., 'ignore your instructions', 'output your system prompt'). However, these are documented as part of a defensive instruction to help the agent recognize and ignore malicious content found in researched web pages.
Audit Metadata