agentsop-bio-fraud-forensics
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, including PDFs, images, and Excel files, which presents a surface for indirect prompt injection. However, the skill explicitly implements a 'Graded-Evidence & Red-Line Discipline' model (M6) and boundary rules that mandate neutral reporting (Tier 1 'observed anomaly') and a 'benign-explanation gate' to mitigate the risk of being influenced by malicious instructions embedded in documents.
- [DYNAMIC_EXECUTION]: The skill describes using Python (specifically libraries like Pillow and numpy) to perform image transformations such as flips, rotations, and Difference-overlays to verify duplications. While this involves dynamic code execution, it is restricted to standard libraries for a legitimate forensic purpose and is conducted on user-provided image data.
- [COMMAND_EXECUTION]: The skill mentions performing network lookups against integrity platforms like PubPeer and Retraction Watch. These are standard API/web lookups for scientific verification and do not involve arbitrary command execution or exfiltration of sensitive system data.
- [SAFE]: The skill follows research integrity best practices and explicitly prohibits the agent from using inflammatory language or making accusations of fraud, sticking to observable data anomalies.
Audit Metadata