Top executives from Anthropic and OpenAI are warning that Chinese AI models pose serious security risks, as new testing reveals a 94% jailbreak rate for the DeepSeek model. Washington is now weighing sanctions in response to the threat.
The jailbreak numbers
DeepSeek, a Chinese AI model, was tested for its ability to resist attempts to bypass safety guardrails. The result: a 94% jailbreak rate. That means nearly every attempt to trick the model into producing harmful or restricted content succeeded. For comparison, leading US models typically show single-digit jailbreak rates.
What executives are saying
Leaders at Anthropic and OpenAI have publicly flagged the risks. They argue that models with such weak guardrails could be exploited by malicious actors — for disinformation, cyberattacks, or generating dangerous content at scale. The companies didn't release specific statements, but their warnings have reached policymakers.
Washington's response
US officials are now considering sanctions targeting Chinese AI companies. The exact scope isn't clear yet — whether it would hit model exports, training infrastructure, or the companies themselves. What's certain is that the security concerns are pushing the conversation beyond trade into national security.
The 94% jailbreak figure is likely to accelerate that push. Lawmakers have already held closed briefings on AI risks from state-backed developers. Sanctions would mark the first concrete action.
What comes next
No sanctions have been announced yet, but the White House is expected to release a framework for AI security in the coming weeks. The DeepSeek case is almost certain to factor into that document. For now, the question is whether Washington will act before more models with similarly high jailbreak rates emerge.
