deep-researcher

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources including PubMed and clinical guideline websites (ACC, ESC, ADA, AHA) using pubmed_fetch_contents and web_fetch tools.
  • Ingestion points: Research data is ingested from external APIs and websites and stored in local files such as sources/pubmed.md (SKILL.md, Step 3).
  • Boundary markers: While explicit delimiters for external content are not mandated, the workflow includes a dedicated validation phase.
  • Capability inventory: The skill employs file system operations (mkdir, cat) and data synthesis tools.
  • Sanitization: The methodology includes a mandatory validation step (Step 4) where the agent is instructed to cross-reference findings and assess source credibility.
  • [COMMAND_EXECUTION]: The skill instructs the agent to use standard shell commands for project organization and context management.
  • Evidence: The workflow uses mkdir -p ~/research_{topic} to establish a project structure and cat to read gathered evidence during the synthesis phase (SKILL.md). These operations are benign and support the research process.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 06:31 AM
Security Audit — agent-trust-hub — deep-researcher