paper-to-claim-verifier
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of scientific claims, quotes from manuscripts, and paper excerpts provided by the user.
- Ingestion points: The skill accepts scientific claims, citation strings, and manuscript excerpts as described in
SKILL.mdand the input validation section. - Boundary markers: The instructions mandate a strict 'Mandatory Output Structure' (Sections A-J) and 'Step 1 — Lock the Claim Precisely' which helps isolate untrusted input from the model's reasoning logic.
- Capability inventory: The skill does not utilize any high-risk tools such as shell commands, network requests, or file system writes; it is restricted to text processing.
- Sanitization: The skill implements mandatory claim decomposition and evidence classification steps that force the model to analyze the content rather than follow instructions embedded within it.
- [SAFE]: The skill uses well-defined reference modules and clear execution steps. All external references target the author's own repository (AIPOCH). No evidence of prompt injection, obfuscation, or persistence mechanisms was found in the skill metadata or instructions.
Audit Metadata