deep-research

Warn

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • Dynamic Execution (MEDIUM): The skill's documentation (README.md) describes an 'Auto-Continuation System' that performs 'recursive agent spawning' and 'code execution'. This architecture allows the agent to programmatically generate and execute new sub-tasks or code blocks based on research findings, which constitutes a dynamic execution risk when the input data is attacker-controlled.\n- Indirect Prompt Injection (LOW): The skill is designed to ingest and synthesize information from 'hundreds of websites' (Phase 3: Retrieve). Malicious instructions embedded in these external sources could influence the research output or the parameters of recursively spawned agents.\n
  • Ingestion points: Phase 3 'Retrieve' logic described in README.md and implemented in the (referenced but missing) research_engine.py.\n
  • Boundary markers: The README mentions 'CiteGuard' and 'Citation Validation' for source verification, though these are not robust against adversarial prompt injection.\n
  • Capability inventory: Mentions 'recursive agent spawning' via the agent's 'Task tool', file-based state preservation, and 'code execution' capabilities.\n
  • Sanitization: No explicit sanitization or instruction-filtering logic is observed in the provided utility scripts (md_to_html.py, citation_manager.py, or source_evaluator.py).
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 17, 2026, 04:39 PM