grounded-research
Pass
Audited by Gen Agent Trust Hub on Mar 28, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists entirely of markdown instructions aimed at improving the accuracy and reliability of the agent's research tasks. No malicious patterns, obfuscation, or unauthorized data access were detected.
- [PROMPT_INJECTION]: The instructions are designed to constrain the agent's behavior to follow Anthropic's hallucination reduction guidelines. There are no attempts to override system prompts or bypass safety filters.
- [INDIRECT_PROMPT_INJECTION]: The skill acknowledges that it will process untrusted external data such as web pages and documents, creating a surface for indirect prompt injection.
- Ingestion points: External documents, web pages, and long-form sources (SKILL.md).
- Boundary markers: The skill mandates the use of specific formatting for extractions (e.g.,
> "exact quote" — [Source]) to clearly distinguish source material from analysis. - Capability inventory: The skill utilizes the agent's existing capabilities for web searching and file reading.
- Sanitization: It implements a defensive 'Extract Before Analysing' protocol, which forces the agent to explicitly identify source quotes before synthesis, acting as a manual verification step against injected instructions in external data.
Audit Metadata