support-with-evidence
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill provides a highly structured and benign instructional framework for an agent to evaluate claims using external web research.
- [SAFE]: It contains extensive internal 'Honesty Rules' and 'Anti-Patterns' that explicitly prevent common AI risks like hallucination, confirmation bias, and source fabrication.
- [INDIRECT_PROMPT_INJECTION]: The skill represents a potential surface for indirect prompt injection as it processes untrusted user claims to generate search queries and analysis.
- Ingestion points: The agent accepts a body of text (arguments, claims, or thesis) at the end of the skill context.
- Boundary markers: There are no explicit technical delimiters (e.g., XML tags or unique markers) surrounding the user input section.
- Capability inventory: The skill is configured to use web search tools and present findings to the user; it lacks high-risk capabilities like file system writes or shell command execution.
- Sanitization: The 'Falsifiability Gate' and 'Claim Extraction' phase act as a semantic filter, as the agent is instructed to discard or list separately any input that is not a falsifiable claim, which would likely include traditional prompt injection imperatives.
Audit Metadata