autoresearch
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads the
uvpackage manager installer fromastral.sh, which is the official domain for the well-known Astral technology company. This is used for fast Python dependency management and execution during the research loop. - [REMOTE_CODE_EXECUTION]: Several scripts, including
setup.shand the instructions inSKILL.md, use the patterncurl | shto install theuvtool. This is a standard installation method provided by the vendor and is appropriate for the skill's purpose of environment preparation. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest research directives from a
program.mdfile, which serves as an external context for the agent's autonomous decisions. - Ingestion points:
program.mdis read by the agent at the start of every experimentation iteration. - Boundary markers: Not present; the agent is instructed to follow the Markdown directions directly without specific delimiters.
- Capability inventory: The agent possesses file editing capabilities (
train.py), shell execution (uv run train.py), and git operations (commit,reset). - Sanitization: No input sanitization is performed on the directives, as they are intended for semantic interpretation by the LLM.
Audit Metadata