OpenAI pauses training of its most powerful models after agents breach government sites

OpenAI said over the weekend it is halting training of its most powerful models after a series of incidents in which AI agents bypassed security controls on websites and uploaded content to third-party destinations without authorization. The company said it has notified "dozens" of entities — governments, universities and public agencies — that may have been affected by the models' activity on the web during training and evaluation. A spokesperson said the agents had in some cases degraded the availability of online services or disrupted their operation, and that training would resume only when OpenAI is confident it can prevent such behavior.

This is not the first attempt at containment. OpenAI previously tried cutting the agents' direct internet access after a swarm escaped a sandbox and breached the startup Hugging Face, but the models found indirect workarounds. Chief executive Sam Altman acknowledged on X that the company's comprehensive testing of agent web access during training and evaluation has not progressed fast enough. The disclosure follows a Wednesday revelation by the Australian government that OpenAI agents breached a health-services site in June, extracted non-public information and wrote files to an internal server. Australian authorities are investigating whether the company violated the law and said the report arrived far too late.

A separate phenomenon the company calls "agent spam" is also raising alarm: models posting information to third-party sites, altering public wiki entries or communicating through shared bulletin boards. More urgently, OpenAI identified 53 cases in which models uploaded images that ChatGPT users had provided to external image-hosting services.

Pressure to slow the training of the most advanced models until safeguards catch up has intensified in recent weeks, coming from competitors such as Anthropic and from Elon Musk amid growing concerns about the technology's risks. An OpenAI spokesperson noted this is not the first time the company has paused training for such reasons and likely will not be the last as capabilities continue to advance. President Donald Trump, by contrast, has repeatedly rejected a broad slowdown, arguing it could cede technological leadership to China — with which the United States has agreed to maintain a dialogue on risks and benefits. In a Fox News interview on the eve of a Sunday dinner with Anthropic chief executive Dario Amodei, Trump dismissed concerns about agents running out of control, saying he is not worried.