copywriting-neta-injection

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

• Third-party content exposure detected (high risk: 0.85). Outsider free text is ingested via the runtime WebSearch-only retrieval in protocols/copy-neta-injection.md → Phase A (Path A-1 SNS/Meme and Path A-2 literary attribution verification), where the agent reads external page/snippet text from URLs returned by site: queries and uses it to build candidate references and canonical forms.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly requires runtime WebSearch retrieval (WebSearch-only rule / Path A-1 and A-2) that queries and pulls content from external sites (e.g., site:twitter.com / site:x.com, site:reddit.com, site:nicovideo.jp, aozora.gr.jp, gutenberg.org, archive.org, ctext.org) and uses those fetched pages to build candidate catalogs and drive the Chain-of-Thought / injection prompts, so these external URLs are runtime dependencies that directly control agent outputs.