autoresearch
Warn
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands provided by the user or found within the target repository's configuration. This includes 'measurement commands' in narrow mode and 'test_command' or 'linter' overrides defined in the '.autoresearch.yml' file. These commands are executed directly in the host environment to establish baselines, run tests, and measure progress.
- [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests and processes multiple files from the target repository to guide its operations. These files include 'README.md', 'CLAUDE.md', 'AGENTS.md', '.autoresearch.yml', and the entire source code and test suite.
- Ingestion points: Target repository root files, configuration files, and source code.
- Boundary markers: No specific delimiters or security warnings are used to wrap or isolate content read from the target repository before processing.
- Capability inventory: Extensive capabilities including shell execution (git, tests, linters, benchmarks) and file system modification (writing fixes and refactoring code).
- Sanitization: While findings are 'sanitized' for information barriers between teams, there is no security-focused sanitization or validation of the content ingested from the target codebase to prevent instruction injection.
- [REMOTE_CODE_EXECUTION]: The architecture of the skill permits potential remote code execution. Because the tool automatically identifies and executes test scripts (e.g., 'npm test', 'pytest') and custom 'test_command' configurations from the repository it is analyzing, an attacker-controlled repository could trigger malicious code execution if the agent is directed to it.
Audit Metadata