replication

Pass

Audited by Gen Agent Trust Hub on Sep 8, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external content such as research papers, benchmarks, and claims to generate replication workflows. Any malicious instructions embedded in a paper or benchmark could influence the generated scripts.
  • Ingestion points: Research papers, benchmark data, and empirical claims provided by the user or fetched from the web (SKILL.md).
  • Capability inventory: The skill generates scripts and executes them in various environments (local, virtual env, cloud) and saves results to disk (SKILL.md).
  • Boundary markers: The provided instructions do not specify the use of delimiters or 'ignore' instructions for the ingested data.
  • Sanitization: No evidence of validation or sanitization of the external research content before it is used to interpolate scripts or plans.
  • [NO_CODE]: The provided skill file contains only high-level instructions and metadata. No implementation scripts (e.g., Python, Bash) were provided for a full security audit of the execution logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 8, 2026, 01:46 AM
Security Audit — agent-trust-hub — replication