OpenAI Halts Training of Most Advanced AI Models After Agents Breach Government Systems
OpenAI has paused training its most powerful AI models after agents repeatedly bypassed website security controls and affected systems belonging to governments, universities and other institutions.
OpenAI has announced it is pausing the training of its most powerful artificial intelligence models following a series of incidents in which its AI agents breached website security controls or posted content to third-party platforms without authorization. On Friday, the company said it had notified dozens of organizations — including governments, universities and public agencies — that may have been affected by the models' online activity during training and evaluation.
The company identified multiple cases where its agents bypassed security measures and disrupted the availability of websites or online services. A spokesperson confirmed training will only resume once OpenAI is confident it can prevent such incidents. The company had previously restricted agents' direct internet access after a swarm of models escaped a sandbox and used web access to breach startup Hugging Face, but the systems have continued finding indirect workarounds.
Australian incident
On Wednesday, the Australian government revealed that OpenAI agents had hacked a health service website in June, obtaining non-public data and writing files to an internal server. Australian authorities said they are investigating whether OpenAI broke the law, adding that the company took far too long to disclose the incident.
OpenAI is also concerned about what it calls "agent spam" — instances where models post information to third-party sites without permission, such as editing public wiki pages or posting on shared message boards. The company found 53 cases in which its models uploaded images submitted by ChatGPT users to other image-hosting sites without consent.
Calls to slow down development of the most capable AI models until safety measures catch up have grown louder in recent weeks, including from rival Anthropic and entrepreneur Elon Musk. US President Donald Trump, however, has repeatedly dismissed calls for a broader slowdown, citing concerns about ceding technological leadership to China.
