fractal

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

• Third-party content exposure detected (medium risk: 0.30). Fractal’s runtime ingests outsider-authored free text provided as the user “question” (or via --question-file) and passes it into multiple engine legs, including a runtime-supplied judge synthesis, via scripts/fractal.py → build_leg_prompt() → engine_legs.call_engine().

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill makes live HTTP calls to DeepSeek at https://api.deepseek.com (scripts/lib/deepseek_router.py), and that classifier + execution API is invoked at runtime to decide routing and to run routed tasks, so this external URL directly controls where prompts are executed.