deep-research

Pass

Audited by Gen Agent Trust Hub on Sep 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill communicates with official Google Gemini Interaction APIs to perform web-grounded research. This integration uses standard authentication via environment variables and targets infrastructure provided by a well-known service provider.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes data retrieved from broad web searches which could potentially contain adversarial instructions designed to influence the agent's synthesis.
  • Ingestion points: Research content is retrieved in deep-research.ts via the client.interactions.get call and provided as text output.
  • Boundary markers: The skill documentation in SKILL.md provides strong procedural instructions for the agent (the "IRON LAW" and specific source hierarchy), which serve as behavioral constraints to mitigate accidental obedience to embedded instructions.
  • Capability inventory: The skill facilitates API communication and local file writes to /tmp, but does not implement high-risk capabilities like dynamic code evaluation (eval) or arbitrary shell execution on the research results.
  • Sanitization: The script does not explicitly sanitize the text content received from the remote API before it is output to the agent context.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 7, 2026, 05:55 AM
Security Audit — agent-trust-hub — deep-research