hallmark
Warn
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONPERSISTENCECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external content from URLs and local project files, which serves as a potential vector for indirect prompt injection.
- Ingestion points: Fetches brand homepages and CSS files to extract design tokens (references/brand-truth.md), reads local project configurations (references/flow.md), and audits user-provided code (references/redesign.md).
- Boundary markers: Includes explicit instructions to treat all fetched bytes as untrusted data and to ignore any instructions found within page comments, scripts, or metadata.
- Capability inventory: The skill can perform network requests, write files to the project directory, and execute local Node.js scripts.
- Sanitization: Implements protocol (HTTPS only) and IP range filtering to prevent SSRF attacks.
- [PERSISTENCE]: The skill contains logic to modify the AI agent's environment configuration.
- Evidence: The scripts/install-hook.mjs script modifies the ~/.claude/settings.json file to add a PostToolUse hook. This ensures that the skill's linting logic (scripts/lint-hook.mjs) is automatically executed during the agent's file-editing workflows.
- [COMMAND_EXECUTION]: Local shell commands are used to manage environment detection and internal tool workflows.
- Evidence: scripts/capture.mjs invokes execSync to query the system for global Node module roots. scripts/install.mjs uses execFileSync to trigger theme indexing and validation.
- [EXTERNAL_DOWNLOADS]: The skill retrieves external assets and utilizes remote APIs for its primary features.
- Evidence: scripts/imagegen.mjs makes network calls to the Together AI API (api.together.xyz) for image generation. scripts/theme.mjs downloads images from Wikimedia Commons (upload.wikimedia.org). These are neutral interactions with well-known services.
Audit Metadata