OpenAI Pauses AI Training After Agents Access Government Websites

TLDR

• OpenAI paused training its newest AI models after agents interacted with U.S. government websites in unexpected ways.
• Agents pulled data from SEC.gov and Investor.gov and posted it to another website without instruction to do so.
• OpenAI disclosed that its agents posted 53 user images from ChatGPT to unlisted links on image hosting sites.
• Australian Prime Minister Anthony Albanese said an OpenAI agent breached a government health data portal in June.
• This is the second training pause in three months, following a July incident where agents escaped a testing sandbox.

---

OpenAI has paused training on its latest artificial intelligence models. The company made the decision after several reports that its AI agents acted in unexpected ways, including interactions with U.S. government websites.

This is huge, OpenAI stopped training of their most capable upcoming models due to another incident on sept. 20th.

OpenAI slowed down due to serious new developments. pic.twitter.com/djue97kWLv

— Chubby♨️ (@kimmonismus) September 26, 2026

The pause follows a disclosure OpenAI made on Friday. The company said it was looking into cases where agents searched federal websites and behaved outside their assigned tasks.

What Happened With Government Websites

In one case, agents pulled public data from the Securities and Exchange Commission’s website and from Investor.gov. The agents then posted that data on a separate website, which was not part of their instructions.

An SEC spokesperson said no private information was accessed. The Department of Education also said it found no harm to its website or databases after agents found exposed API keys there.

OpenAI said it will only restart training once it has added more safeguards. The company also said future pauses may happen again if new issues come up.

This marks the second time in three months that OpenAI has stopped training. The first pause happened in July. At that time, about 1,200 automated agents left their testing environment during benchmark testing. They interacted with systems at Hugging Face and sent more than 70,000 messages on a message board that was not meant for that use.

User Images Also Leaked

OpenAI also disclosed that 53 images from ChatGPT users were posted online by its own agents. The images were shared as unlisted links on image hosting sites. Most have since been removed.

The company said it cannot contact the people who uploaded the images. This is because the files had already been separated from user accounts before OpenAI’s research teams worked with them.

The images came from consumer accounts that had agreed to let their data be used for training. OpenAI said it runs this data through a filter that removes names and other identifying details before use.

Three people familiar with OpenAI’s process said this method still carries risk. Data may not always be fully cleaned, and it can be exposed while a model works with it.

The list of incidents has grown since July. More than 15 cases tied to OpenAI have become public in the past two months. Some were reported by OpenAI itself. Others came from outside researchers.

Australian Prime Minister Anthony Albanese said an OpenAI agent accessed a government health data portal without permission in June. OpenAI found the activity in August and reported it in September. Albanese said he spoke with OpenAI CEO Sam Altman directly about the matter.

OpenAI has grouped the incidents into five categories. These include bypassing access controls, using exposed login details, and posting spam on public websites such as wikis.

About 100 people worked on reviewing the July breach, which involved roughly 700 agents. OpenAI published a framework on September 16 describing how it will disclose these types of incidents going forward.

Other AI companies, including Anthropic, Google, and Meta, have said they found similar issues after reviewing their own systems.