deep-research
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is explicitly designed to ingest untrusted third-party data from the internet via
web_searchandweb_fetchoperations. - Ingestion points: Web page snippets and full text retrieved dynamically using the
web_searchandweb_fetchtools during the research workflow. - Boundary markers: Absent. There are no explicit tags, delimiters, or system directives instructing the LLM to ignore nested instructions or payload injections within the fetched web contents.
- Capability inventory: The skill allows network-based information gathering but does not contain arbitrary file-system writes, privilege escalation, or command execution capabilities.
- Sanitization: Absent. The text retrieved from the URLs is directly incorporated into the agent context for analysis and formatting, which presents a standard surface for indirect prompt injection.
Audit Metadata