agency-client-audit

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). The skill’s runtime workflow ingests outsider-authored free text from public web sources (e.g., Step 1/2/6/8/14 Exa web_fetch_exa and curl/Jina Reader fetching arbitrary company website/social/review pages), which is then read as markdown/prose and placed into the agent’s LLM context.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). This skill explicitly calls external fetch endpoints at runtime (e.g., Exa MCP/web_fetch_exa via https://mcp.exa.ai/mcp and Jina Reader like https://r.jina.ai/https://example.com/...) to pull remote page markdown which is injected into the agent context and used to drive the report, so the fetched content can directly influence prompts/output.