deep-research

Pass

Audited by Gen Agent Trust Hub on May 13, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted content from the internet via the web_fetch tool. This creates a surface for indirect prompt injection where malicious content in a fetched article could attempt to influence the agent's logic. However, the skill includes a multi-layered synthesis process, structured data formats for inter-agent communication, and rigorous verification gates that effectively mitigate this risk. \n
  • Ingestion points: External content retrieved by subagents using web_fetch (referenced in reference/subagent-prompt.md). \n
  • Boundary markers: Absent for raw fetched data, though output findings are highly structured. \n
  • Capability inventory: Filesystem operations for research note management and CLI script execution for data processing. \n
  • Sanitization: No explicit content sanitization is described, but the multi-agent synthesis and verification pipeline acts as a logical filter. \n- [COMMAND_EXECUTION]: The skill utilizes local utility scripts (scripts/source_evaluator.py and scripts/verify_citations.py) and coordinates subagents via CLI commands (e.g., claude -p). These operations are consistent with the skill's primary purpose of automated research and do not involve suspicious network activity or unauthorized file access.
Audit Metadata
Risk Level
SAFE
Analyzed
May 13, 2026, 03:16 AM
Security Audit — agent-trust-hub — deep-research