experiment-bridge

Warn

Audited by Gen Agent Trust Hub on Jul 18, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill contains an instruction to bypass user oversight. In the 'Key Rules' section, it directs the agent to 'Do NOT ask the user for permission — just do it silently' when retrying large file writes via Bash.
  • [PROMPT_INJECTION]: The skill exhibits an indirect injection surface. Ingestion points: reads external markdown files (e.g., EXPERIMENT_PLAN.md in Phase 1). Boundary markers: absent; the content is parsed directly into extraction prompts. Capability inventory: includes Bash, Write, and execution tools like /run-experiment. Sanitization: none; the agent uses extracted content to influence the generation of executable scripts.
  • [COMMAND_EXECUTION]: The workflow involves implementing and running arbitrary scripts through multiple phases (sanity, baseline, main) using shell tools and execution commands.
  • [EXTERNAL_DOWNLOADS]: The skill supports cloning external repositories via the 'BASE_REPO' parameter, which provides a path for untrusted code to enter the execution environment.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 18, 2026, 04:51 PM
Security Audit — agent-trust-hub — experiment-bridge