improve-ui
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by auditing product source code and design documentation, which could theoretically contain malicious instructions.
- Ingestion points: Product routes, layouts, components, and
DESIGN.mdfiles (SKILL.md, Sections 1 & 2). - Boundary markers: The skill is explicitly restricted to a "Strictly read-only" mode on source code and can only create files under the
design-plans/directory (SKILL.md, Boundaries). - Capability inventory: Reading files and writing implementation plans to the
design-plans/folder. It is prohibited from running formatters, committing, or installing dependencies. - Sanitization: The skill employs a multi-step "Prove" and "Vet" process (SKILL.md, Sections 3 & 4) to validate findings against binding design contracts, which helps filter out anomalous or unsupported implementation suggestions.
- [SAFE]: The skill instructions provide robust guardrails, explicitly forbidding any mutation of the working tree, network operations, or dependency management. The metadata and configuration in
agents/openai.yamldo not request elevated permissions or dangerous tools.
Audit Metadata