Model Theft

Model theft refers to the unauthorized copying or extraction of a trained AI model, either by stealing the model file directly or by querying a deployed system repeatedly until its behavior can be reproduced as a near-copy. For enterprises, this puts years of development spend, proprietary training data, and competitive advantage at risk, and it can also expose sensitive information the model learned from. Governance teams should treat model access controls, API rate limits, and monitoring for suspicious query patterns as core protections against it.

1 item

ResearchUS2026-10-26

Compaction Smuggling Flaw Lets Attackers Extract and Copy Frontier AI Reasoning

Security firm Xint disclosed a flaw in OpenAI's context compaction process called Compaction Smuggling. The flaw allowed internal reasoning steps of advanced AI models to be extracted and transferred to weaker ones. OpenAI confirmed the vulnerability, deployed fixes, and paid a $6,000 bug bounty on October 23, 2026. CISA separately warned that Chinese AI developers, including DeepSeek, have been extracting billions of AI outputs from US frontier models since at least late 2024.

OpenAI DeepSeek model theft API security nation-state risk AI distillation CISA intellectual property