autoresearch
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to create and execute a local bash script named
autoresearch.shto run benchmarks. It also optionally executes a backup script located at./scripts/backup-state.sh. - [DYNAMIC_EXECUTION]: The agent writes the logic for
autoresearch.shdynamically based on the goals and constraints inferred from user instructions, then executes the resulting script usingbash -c. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local environment which could contain malicious instructions designed to subvert the agent's autonomous loop.
- Ingestion points: Reading local source files,
autoresearch.ideas.md,experiments/worklog.md, andautoresearch.jsonlduring setup and resumption. - Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded prompts within the ingested source or ideas files.
- Capability inventory: The agent has the ability to write and execute shell scripts, perform git operations (commit, checkout, clean), and modify the file system.
- Sanitization: There is no evidence of sanitization or filtering applied to the content of the project files or the ideas backlog before they are processed by the agent.
Audit Metadata