web-research
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the web via the
WebSearchtool, creating a surface for indirect prompt injection where malicious content on websites could attempt to influence agent behavior.\n - Ingestion points: Web search results (titles, URLs, snippets) are processed in Step 2 and Step 3 of the research protocol in
SKILL.md.\n - Boundary markers: The skill explicitly requires cross-validation of all claims using at least two independent sources and mandates a confidence declaration (95/80/60/<60%).\n
- Capability inventory: The agent is authorized to use
WebSearch, write to the local file system (tmp/research/,docs/knowledge/,tmp/reports/), and spawn subagents viaspawn_agent.\n - Sanitization: The protocol includes a mandatory triage step (Step 3) that classifies sources into tiers (1-4) and deduplicates content to ensure only high-signal information is used.\n- [DYNAMIC_EXECUTION]: The skill conditionally loads and executes additional instructions from local files based on user-provided flags.\n
- Evidence: When the
--chain=deep-diveflag is passed, the skill reads and follows procedures defined inreferences/research-chain.mdand.claude/skills/source-deep-dive/SKILL.md.
Audit Metadata