last30days-glim

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

• Third-party content exposure detected (high risk: 0.95). Outsider free text from public web/social sources (e.g., Reddit comments, X posts, YouTube transcripts, TikTok captions, GitHub issue/PR bodies) is fetched at runtime and then ingested into the LLM judge via glim inference calls in the reranking step (untrusted-content candidates are included in the prompt context).

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

• Potentially malicious external URL detected (high risk: 0.80). The skill makes runtime LLM/inference calls to glim (POST https://surf.cascade.fyi/api/v2/inference/v1/chat/completions) to run the planner and relevance/fun judges whose outputs directly control the agent's query plan and reranking logic, and the pipeline relies on those responses for full-quality runs.