autoresearch
Pass
Audited by Gen Agent Trust Hub on Apr 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the installation script for the well-known 'uv' package manager from astral.sh, a standard service for Python tooling. Evidence: Found in SKILL.md and scripts/setup.sh.\n- [REMOTE_CODE_EXECUTION]: Executes the downloaded 'uv' installer via a shell pipe. This is the official and documented installation method for this well-known tool. Evidence: curl commands in SKILL.md and scripts/setup.sh.\n- [COMMAND_EXECUTION]: Uses local scripts to orchestrate the autonomous research loop, including training execution and git-based model ratcheting. Evidence: The scripts/ directory contains various tools for managing the experimentation workflow.\n- [PROMPT_INJECTION]: The system processes directives from a researcher-authored program.md file to guide the agent's code modifications. Ingestion points: program.md (referenced in SKILL.md and architecture guides). Boundary markers: None. Capability inventory: Modifies train.py and executes training runs using uv. Sanitization: None.
Audit Metadata