deep-research
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The workflow retrieves data from the internet via WebSearch and WebFetch, making it a target for instructions embedded in web content.\n
- Ingestion points: Untrusted content enters the agent's environment through search result snippets and page content in scripts/workflow-script.js.\n
- Boundary markers: The skill uses markdown headers but lacks explicit safety delimiters (e.g., XML tags or specific 'ignore' instructions) around data from external sources.\n
- Capability inventory: Capabilities are limited to WebSearch and WebFetch tools; the skill does not exhibit file-writing or shell-execution behaviors.\n
- Sanitization: The script includes a LABEL_STRIP mechanism to filter terminal control sequences, Unicode bidirectional overrides, and zero-width characters from labels. It also uses structured JSON schemas for validation.
Audit Metadata