plan-eng-review

Warn

Audited by Gen Agent Trust Hub on Aug 7, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The instructions contain directives to override standard agent behavior, stating that 'the skill wins' in the event of a conflict with other system nudges or instructions.
  • [COMMAND_EXECUTION]: The skill's preamble and various workflows execute multiple local bash commands and specialized binaries (such as gstack-config, gstack-slug, and gstack-repo-mode) to manage session data and project configurations.
  • [REMOTE_CODE_EXECUTION]: Through the 'Outside Voice' feature, the skill can invoke the codex exec tool, which executes code locally that has been generated by a remote model based on a provided prompt.
  • [COMMAND_EXECUTION]: Dynamic execution patterns are employed, including the use of eval "$(...)" and process substitution (source <(...)) to incorporate the runtime output of local scripts into the agent's environment.
  • [EXTERNAL_DOWNLOADS]: The skill automates git fetch and git merge operations to synchronize artifacts with a remote repository and includes references to external documentation sites like 'garryslist.org'.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 7, 2026, 08:36 PM
Security Audit — agent-trust-hub — plan-eng-review