storm-research

Pass

Audited by Gen Agent Trust Hub on Aug 19, 2026

Risk Level: SAFEPROMPT_INJECTIONNO_CODE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The research workflow is designed to ingest and analyze potentially untrusted data from external URLs, datasets, and files.\n
  • Ingestion points: SKILL.md (Step 2: 'Establish baseline facts') fetches external sources; SKILL.md (Step 4: 'Dispatch one isolated executor per lens') collects raw returns from child processes.\n
  • Boundary markers: Explicit defensive instructions are included in SKILL.md (Step 2) and references/lens-charter.md to treat external text as untrusted and ignore embedded instructions.\n
  • Capability inventory: The skill framing include access to URLs, file paths, and repository names for retrieval purposes.\n
  • Sanitization: The agent is instructed to 'extract evidence' rather than executing or following content from external sources.\n- [SAFE]: The skill implements context isolation (Step 4) to prevent cross-contamination and data leaks between different analytical perspectives (lenses). This represents a security best practice for multi-agent workflows processing external information.\n- [SAFE]: Explicit verification of context isolation is required in the workflow, reducing the risk of accidental prompt contamination from inherited conversations or sibling tasks.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 19, 2026, 01:14 AM
Security Audit — agent-trust-hub — storm-research