eval

Warn

Audited by Gen Agent Trust Hub on Mar 17, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to run a user-defined evaluation command ({eval_cmd}) within the worktree of an agent. This pattern allows for the execution of arbitrary shell commands if the configuration is sourced from an untrusted or improperly sanitized origin.
  • [PROMPT_INJECTION]: The skill implements an 'LLM Judge' mode that consumes and processes agent-generated data, creating a surface for indirect prompt injection.
  • Ingestion points: The judge reads agent result posts located at .agenthub/board/results/agent-{i}-result.md and the output of git diff commands.
  • Boundary markers: Absent. The description does not mention the use of delimiters or 'ignore' instructions to prevent the LLM from obeying commands embedded in agent results.
  • Capability inventory: The skill executes local Python scripts (result_ranker.py, session_manager.py) to rank results and update session states, which could be influenced by a successful injection.
  • Sanitization: Absent. There is no indication that agent-generated content is sanitized, escaped, or validated before being passed to the LLM judge.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 17, 2026, 03:53 PM