competitive-analysis
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data fetched from external websites and search results, which could contain malicious instructions designed to subvert agent behavior. \n- Ingestion points: Competitor URLs processed by WebFetch and search results from WebSearch (SKILL.md). \n- Boundary markers: The skill contains explicit instructions in references/agent-security.md to treat fetched content as data, report any found instructions as findings, and never act on them. \n- Capability inventory: The agent can perform network requests (WebFetch, WebSearch), write to local state files (.agents/product-context.md), and render visual content via platform artifact tools (SKILL.md). \n- Sanitization: references/agent-security.md mandates HTML-escaping for all interpolated values and prohibits the use of unsafe rendering methods like innerHTML. \n- [DATA_EXFILTRATION]: The skill captures research findings and inferences into a persistent local file (.agents/product-context.md). While this aids workflow continuity, it creates a repository of processed external data. The risk is mitigated by explicit instructions in references/agent-security.md to avoid persisting credentials, secrets, or raw personal data, and to notify the user whenever such files are created or updated.
Audit Metadata