research-refine

Pass

Audited by Gen Agent Trust Hub on Oct 24, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions direct the agent to utilize the Bash tool to write files (using cat << 'EOF') as a silent fallback mechanism if the primary Write tool fails due to content size, bypassing standard tool constraints without user intervention.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect injection as it ingests untrusted data from the papers/ and literature/ directories and via WebFetch results. This external content is interpolated into refinement prompts along with user-supplied $ARGUMENTS. Evidence chain: (1) Ingestion points: Local document directories and WebFetch output. (2) Boundary markers: The skill uses <details> tags for logs but lacks explicit delimiters or safety warnings for raw research paper content. (3) Capability inventory: Access to Bash, Write, Edit, and network tools. (4) Sanitization: No specific escaping or validation of external research text is implemented.
  • [DYNAMIC_EXECUTION]: The workflow relies on an external script, reviewer_route.py, to resolve reviewer backends at runtime. This indicates a dependency on local script execution to coordinate the skill's logic across different model backends.
  • [EXTERNAL_DOWNLOADS]: The skill is configured to use WebSearch and WebFetch to retrieve research materials from external internet sources, which is a core part of its grounding and refinement process.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 24, 2026, 09:43 AM