autoresearch

Fail

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PRIVILEGE_ESCALATION]: The runner_template.py script invokes claude -p with the --dangerously-skip-permissions flag, bypassing user confirmation for arbitrary actions such as file modifications and shell command execution.\n- [DYNAMIC_EXECUTION]: Instructions for an inner agent are dynamically assembled at runtime based on repository state and executed via claude -p. This process incorporates potentially untrusted data into the execution logic.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from program.md, results.tsv, and git history to generate agent prompts without sanitization or boundary markers.\n
  • Ingestion points: scripts/runner_template.py reads repository metadata, logs, and user-provided markdown files.\n
  • Boundary markers: The skill uses simple string interpolation to build prompts without robust delimiters or safety instructions for the processed data.\n
  • Capability inventory: The autonomous loop has the ability to run build commands, test suites, and git operations via subprocess.run, and arbitrary agentic tasks via claude -p.\n
  • Sanitization: No sanitization or validation of the ingested repository content is performed before it is passed to the LLM for hypothesis generation.\n- [COMMAND_EXECUTION]: Extensive use of subprocess.run across multiple scripts to execute shell commands for git operations, build tools, and evaluation gates. Scripts involved include runner_template.py, scaffold.py, eval_gen.py, and eval_base.py.\n- [EXTERNAL_DOWNLOADS]: The scripts/eval_gen.py script attempts to communicate with api.exa.ai via a POST request to fetch testing patterns if an API key is provided in the environment.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 26, 2026, 06:34 PM