higgsfield

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.85). The skill contains powerful automation that can modify and commit its own code, create scheduled cron tasks, run arbitrary shell commands, fetch and write arbitrary external content to local files, and drive browser uploads/downloads — capabilities that, while presented for orchestration, provide clear primitives for persistence, remote control, and data exfiltration if abused or if LLM/subagent outputs are compromised.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 1.00). The skill's Phase 2.6 "Visual research" step explicitly dispatches visual-researcher agents that download candidate reference images and write reference_urls_start/end and research_notes into claims (saved under $OUTPUT_DIR/references/claim_/), and those researcher-sourced reference images and URLs are then read by the Creative Director, Shot Planner, image-worker, and image-reviewer and can be attached to generation prompts—i.e., the agent fetches and ingests arbitrary third-party web images/URLs and uses them to drive prompt composition and submission.