paper-claim-audit

Pass

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill exhibits a potential vulnerability to indirect prompt injection because it instructs a reviewer agent (GPT-6-Astra) to ingest and process untrusted external data (LaTeX source files and raw result data) without explicit sanitization or strict boundary markers.
  • Ingestion points: The skill reads and processes all .tex files in the paper directory and all .json, .csv, .tsv, and .yaml files in the results and outputs directories.
  • Boundary markers: The prompt template provided for the reviewer agent in Step 2 does not include unique delimiters or specific 'ignore embedded instructions' warnings to prevent the agent from following instructions that might be maliciously embedded in the research paper text or data files.
  • Capability inventory: The reviewer agent has access to analysis capabilities, and the overall skill has permissions for Bash(*), Write, Edit, and spawn_agent (via the executor).
  • Sanitization: There is no evidence of content filtering, escaping, or validation of the text extracted from paper files before it is provided to the reviewer agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 10, 2026, 09:36 AM