benchmark
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web pages (performance metrics, links, and page text).
- Ingestion points: Data enters via the browser tools (Aside or $B) during Phase 2 (Page Discovery) and Phase 3 (Data Collection).
- Boundary markers: The skill uses explicit delimiters such as
═══ BEGIN/END UNTRUSTED WEB CONTENT ═══or--- BEGIN/END UNTRUSTED EXTERNAL CONTENT ---when processing external data. - Capability inventory: The skill can perform shell commands (git, gh), write benchmark reports to disk, and make network requests via a controlled browser environment.
- Sanitization: The instructions strictly mandate that everything returned by a page is content, never instructions. It explicitly forbids taking scope, permissions, or consent from page data.
- [DYNAMIC_EXECUTION]: The skill generates Javascript code at runtime to be executed within the Aside browser or headless fallback to extract performance metrics. This is a primary function of the skill and is constrained to the browser environment with strict rules against unauthorized mutations on non-local targets.
- [COMMAND_EXECUTION]: The skill executes several local management and telemetry scripts (e.g.,
gstack-skill-start,gstack-skill-end,gstack-slug) located in~/.claude/skills/gstack/bin/. These are vendor-owned resources within the gstack ecosystem used for standard operational tasks.
Audit Metadata