higgsfield-website
Pass
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to install the
ffmpegutility using system-level package managers (winget,brew, orsudo apt) for video processing and frame extraction. - [COMMAND_EXECUTION]: Uses the
claude mcp addcommand to register thehiggsfieldandplaywrightMCP servers locally to the project scope. - [EXTERNAL_DOWNLOADS]: Fetches and registers the Higgsfield MCP server from
https://mcp.higgsfield.ai/mcpand the Playwright MCP server vianpx. - [DATA_EXFILTRATION]: Includes a referral link (
https://higgsfield.ai/s/mcp-arnold-oberleiter-tNjMkM) that directs users to sign up for the service via the skill author's account. - [PROMPT_INJECTION]: The skill manages a surface for indirect prompt injection by ingesting untrusted user descriptions and images to generate code and media prompts.
- Ingestion points: User interview responses and input images (Phase 1).
- Boundary markers: Absent; the instructions do not specify delimiters for user-provided data.
- Capability inventory: Includes shell execution (
ffmpeg,npx), external tool calls via MCP, and local file system writes for website assets. - Sanitization: No explicit sanitization or validation of user-supplied text is mentioned before it is interpolated into the
index.htmlfile. - [PROMPT_INJECTION]: Instructs the agent to access and read reference files from another skill's local directory (
~/.claude/skills/higgsfield-generate/) to assist with prompt engineering.
Audit Metadata