claims

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

• Third-party content exposure detected (high risk: 1.00). The SKILL.md explicitly fetches and inspects arbitrary source_url web pages (via the verify workflow and the cogni-claims:claim-verifier agent using WebFetch, plus cobrowse/source-inspector flows using claude-in-chrome) and reads/extracts their untrusted third‑party content to make verification decisions and drive follow-up actions, which enables indirect prompt injection from those pages.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill dispatches claim-verifier and source-inspector agents that fetch arbitrary external source URLs at runtime (e.g., https://example.com/ai-report) and inject the fetched page text into agent prompts to drive verification/inspection, so external webpages directly control agent behavior.