research-conventions
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides instructions for processing data from untrusted external sources, such as web searches and regulatory filings, creating a potential attack surface.
- Ingestion points: As specified in
references/evidence.md, data is ingested through several tools includingget_sec_filing,WebFetch, andWebSearch. - Boundary markers: The skill does not define structural boundary markers (like specific delimiters), but it mandates a logical labeling system (e.g.,
fact,company claim,street estimate) to help the agent track the provenance of every claim. - Capability inventory: The skill governs the creation of research artifacts (prose, slide decks, and workbooks) but does not include scripts that execute arbitrary commands, write to sensitive system files, or perform network exfiltration.
- Sanitization: The skill enforces a strict "Source tiers by claim family" hierarchy and a mandatory "conflict register" to verify and de-conflict data, prioritizing primary documents over third-party summaries, which acts as a verification layer against potentially poisoned data.
Audit Metadata