experiment-audit
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONINGDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the user's project directory, creating a significant attack surface for indirect prompt injection.
- Ingestion points: The skill scans and reads evaluation scripts, result files, experiment trackers, and paper claims directly from the project directory.
- Boundary markers: The skill does not implement delimiters or provide 'ignore embedded instructions' warnings when passing file content to the reviewer agent.
- Capability inventory: The agent has access to powerful tools including Bash, Read, Write, Edit, and various external MCP tools for model communication.
- Sanitization: No evidence of sanitization or filtering of the content read from project files is present before it is processed.
- [METADATA_POISONING]: The skill's description and configuration constants reference non-existent or hypothetical AI models such as 'GPT-6-Astra' and 'GPT-5.5 Pro'. This metadata is deceptive regarding the skill's actual technical capabilities and the specific models used for the integrity audit.
- [DYNAMIC_EXECUTION]: The skill utilizes the spawn_agent tool to dynamically generate and execute instructions for a sub-agent at runtime. This process involves constructing a prompt from file paths and project data gathered during execution, which could be manipulated if the project structure itself contains malicious identifiers.
Audit Metadata