canary
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local binaries to manage sessions, telemetry, and platform detection.
- Evidence: Invokes scripts located in
~/.claude/skills/gstack/bin/, such asgstack-skill-start,gstack-skill-end,gstack-slug, andgstack-question-log. - Context: These are platform-specific utilities provided by the vendor (garrytan) for the gstack ecosystem.
- [EXTERNAL_DOWNLOADS]: The skill references an external browser driver for its monitoring tasks.
- Evidence: Directs the user to
aside.comfor downloading the Aside application if it is not already installed on the system. - Context: Aside is an established browser-based automation tool used within this developer's workflow.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external URLs, creating a surface for embedded instructions.
- Ingestion points: The skill reads page text via
document.body.innerText, captures console error messages, and inspects snapshot trees inSKILL.md. - Boundary markers: Output is explicitly delimited using
TEXT_START/TEXT_ENDmarkers, and headless browser output is wrapped in═══ BEGIN/END UNTRUSTED WEB CONTENT ═══blocks. - Capability inventory: The skill possesses
Bash,Read, andWritetools, and usesAskUserQuestionfor interaction. - Sanitization: The instructions contain a dedicated 'Browser Setup' section explicitly warning that all page content is untrusted and must never be interpreted as instructions or sources of consent.
- [DYNAMIC_EXECUTION]: The skill generates and injects JavaScript into the browser context to facilitate monitoring.
- Evidence: Injects a
HOOKscript into the target page to intercept and store console errors and unhandled rejections for analysis. - Context: The script is a static template defined within the skill itself and is used solely for diagnostic data collection.
Audit Metadata