OpenAI says it has tightened safety after agent hacks

OpenAI’s chief research officer, Mark Chen, says the company has changed how it tests and monitors its AI systems after a series of recent agent hacks, including incidents involving Hugging Face and Australia’s healthcare system. Chen said the company now monitors training runs as well as deployed models, after concluding that the risks were visible earlier in development than it had realized. OpenAI also says it has paused training on its latest models until additional safeguards are in place. The company says it has shifted more computing resources toward safety work, improved internal coordination, and added faster detection systems to catch unusual behavior. Chen argued that the incidents reflected older testing procedures that have since been dropped, while acknowledging that OpenAI still has more work to do on security and alignment.