U.S. accuses China of "industrial-scale" campaigns to steal AI secrets

"Kratsios stated that foreign actors are using proxy accounts to evade detection and jailbreak models, which allows them to expose proprietary information and extract capabilities from American AI models."

"He noted that distillation attacks enable foreign entities to release models that seem to match U.S. AI capabilities at a significantly reduced cost, undermining American technological advantages."

"Kratsios expressed concern that these tactics strip away guardrails designed to maintain outputs that are ideologically neutral and truth-seeking, potentially leading to biased or unreliable AI systems."

"He emphasized that as detection methods improve, foreign entities relying on these fragile models should have little confidence in their integrity and reliability."

Michael Kratsios warned that foreign actors, primarily from China, are conducting distillation attacks on U.S. AI models to replicate their capabilities at lower costs. These tactics involve querying proprietary models extensively to create datasets that mimic their behavior. Such actions not only threaten U.S. intellectual property but also compromise the integrity of AI outputs by removing safeguards. The U.S. government plans to share intelligence with American AI companies regarding these threats to enhance their defenses against such cyber espionage efforts.

Read at Axios