Model theft targets extraction of model weights, behavior, or proprietary capabilities through repeated querying, endpoint abuse, or supply chain compromise. Attackers can replicate functionality, bypass business controls, and monetize stolen model value. Theft also enables adversaries to reverse engineer defenses and craft tailored attacks.
Protect model endpoints with robust authentication, abuse detection, and query-pattern monitoring for extraction attempts. Limit high-fidelity outputs where appropriate and apply watermarking or fingerprinting strategies. Harden deployment pipelines and secrets management to reduce model artifact exposure.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2024Latest
#10llm10-2024 Model Theft
