review-fix-loop

Fail

Audited by Gen Agent Trust Hub on Apr 1, 2026

Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill automatically identifies and executes shell commands defined in project configuration files during the verification step.
  • Evidence: Step 8 (Verify) searches for and runs lint and test commands from package.json (scripts.lint/test), pyproject.toml, Makefile, and Justfile.
  • Risk: A malicious repository could define these standard configuration keys to execute unauthorized code on the host system when the agent attempts to verify fixes.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection where malicious instructions from analysis tool outputs are passed to subagents.
  • Ingestion points: Step 4 parses fix_instructions from the stdout and stderr of tool execution.
  • Boundary markers: Absent. Tool output is captured and parsed without sanitization or delimiters before being sent to subagents.
  • Capability inventory: Subagents dispatched in Step 7 have the capability to modify the codebase.
  • Sanitization: Absent. Extracted fix instructions are interpolated directly into the subagent prompt.
  • [COMMAND_EXECUTION]: User-provided arguments are used to construct shell commands without sufficient sanitization.
  • Evidence: Step 1 and Step 3 use $ARGUMENTS directly as the command to be executed in the shell.
  • Evidence: Step 2 extracts the first word of the command and interpolates it into a bash command: command -v .
  • Risk: If the input contains shell metacharacters (e.g., semicolons or pipes), it could lead to arbitrary command execution.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Apr 1, 2026, 01:32 PM