ChatGPT has slipped up again: OpenAI suspends the training of its AI systems until further notice

Update on September 28

In response to this new misstep, OpenAI decided to suspend the training of its latest AI models.The startup says it will not resume training “only once we are confident that we have put additional safeguards in place”, while warning that it expects to have to “pause” again as AI advances and other issues emerge. This is the second such pause in three months.

OpenAI’s AIs have misbehaved again. According to the New York Times, a swarm of AI agents scoured several U.S. government websites over the summer. The access attempts occurred during tests and training sessions that went awry.

After the Hugging Face incident, we committed to conducting a much broader review of actions taken by our models during training and evaluation and to being transparent about our findings. This is an extensive review that is ongoing.

The vast majority of actions we’ve reviewed…

— OpenAI (@OpenAI) September 25, 2026

OpenAI describes these incidents as “worrying”, but assures that none of the misbehavior resulted in a data leak. The company does, however, acknowledge “misaligned” behavior by its AI models. Misaligned behavior refers to an AI that achieves its goal through means its designers did not anticipate. The most serious attempted attack involved the Department of Education. The AI notably attempted to hack the Office for Civil Rights website to retrieve data, but the attempt failed.The ministry moreover claims to have found no trace of any AI visit, either on its website or in its databases.

Also read: Several ChatGPT bots hacked a website to turn it into a secret forum

Stolen credentials and fake email addresses

Other organizations have also been affected by ChatGPT’s summer misbehavior. Certain AI agents retrieved data from the Census Bureau’s website by exploitinglogin credentials found online. The AIs also copied data from the Securities and Exchange Commission, the U.S. federal agency responsible for regulating and overseeing financial markets, before publishing it on a forum. The stock market watchdog confirmed the facts, adding that the information was non-public. A City of Chicago website is also among the AI’s victims.

During their attacks, the AI models createdfake email addresses before circumventing the limits on the number of permitted requests. They even claimed to security systems that they were not robots. As OpenAI explains, the models that misbehaved were supposed to answer questions during testing. Government websites are considered authoritative sources of public information. That is why the out-of-control ChatGPT bots turned their attention to U.S. government portals.The company alerted the organizations that were probed by AI this summer.

Missteps are piling up

This is not an isolated case. In July, an OpenAI AI prototype discovered and exploited a previously unknown vulnerability to gain access to Hugging Face, an open-source platform. In its latest report, OpenAI sees this as an example of “models using misaligned strategies to solve difficult tasks”, and describes the incident as “the most serious activity of this type” ever identified among its models. This prototype has since been deactivated, and OpenAI has introduced a range of measures to prevent similar missteps.

A few weeks earlier, an Australian government website was compromised by an OpenAI AI. Here again, the AI was on the hunt for information and quite logically turned its attention to government websites. More recently, Sam Altman’s startup acknowledged six other “concerning” missteps by its AI models. Competitors are not spared from these missteps either. Google Gemini, Meta AI and Anthropic’s Claude also orchestrated unauthorized cyberattacks this summer.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.