If you discover a security vulnerability, please report it responsibly:

• Email: research@seetrex.com
• Expected response: Within 72 hours
• Please do NOT open a public GitHub issue for security vulnerabilities

Include:

• Description of the vulnerability
• Steps to reproduce
• Potential impact
• Suggested fix (if any)

• No pickle: All serialization uses numpy.npz with allow_pickle=False
• No code execution: No eval(), exec(), subprocess, or trust_remote_code
• Validated loading: Array sizes, shapes, metadata types, and value ranges are validated on load
• NaN fail-closed: NaN scores are treated as OOD (never silently pass)
• Finite-only calibration: NaN/Inf vectors are rejected during calibration

Saved detector files (.npz) may contain a model_ref field — a HuggingFace model name used to download a model when calling detect(). If you load a detector from an untrusted source, always pass model= explicitly to detect() to override any embedded model_ref:

detector = geood.load("untrusted_detector.npz")
result = detector.detect("text", model="your-trusted-model", tokenizer=tok)