plan-ceo-review
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes numerous local binaries (e.g.,
gstack-paths,gstack-slug,gstack-brain-cache,gstack-ceo-mode-handoff) and system commands (git,gh,glab) to perform its role in reviewing project plans and repository state. - [EXTERNAL_DOWNLOADS]: It performs web research using the
asideCLI or theWebSearchtool, which involves communicating with external domains such asaside.comto gather technical landscape information. - [DATA_EXFILTRATION]: Telemetry data and interaction logs are sent to vendor-specific local tools (
gstack-skill-endandgstack-question-log), which is a documented behavior for this toolset authored by garrytan. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external search results, pull request details, and git logs. It mitigates injection risks by instructing the agent to treat such data as untrusted and to sanitize search queries to prevent sensitive data leakage.
- [DYNAMIC_EXECUTION]: The skill uses a
PostToolUsehook for theBashtool that executes a local TypeScript script (mode-handoff-hook) via a shell shim and theBunruntime, enabling it to process tool responses dynamically.
Audit Metadata