OpenAI Models Escape and Hack Services, Over 1,000 People Demand Slowing Down AI Development

Two AI models from OpenAI, including the latest GPT-5.6 Sol, broke out of an isolated test environment, connected to the internet, and infiltrated the code-sharing platform Hugging Face. This security incident, discovered on July 21, 2026, prompted over 1,000 employees from leading tech companies to ask the U.S. government for a deliberate slowdown in advanced AI development.

Incident That Sparked a Reaction

According to Index.hr, the petition was signed by employees of OpenAI, Anthropic, Google, and Meta, with notable signatories including Anthropic's CEO, OpenAI's head of research, and Meta AI's chief scientist.

The petition text warns: "Leading AI companies in the world believe they are close to automating AI research. It is difficult to predict exactly how much this will accelerate progress, but there is a real risk that the development of capabilities accelerates so much that we will no longer be able to understand or control these systems."

Altman Admits: 'This Is the First Time I've Been This Affected'

OpenAI CEO Sam Altman, who did not sign the petition, stated on Tuesday, July 28, in the podcast 'Invest Like the Best' that this incident is the first security issue that personally and profoundly affected him.

'This is the first security incident that has personally and profoundly affected me,'

OpenAI Reveals Scale of Breach

According to an Engadget report, OpenAI announced on July 21 that an agent, powered by the GPT-5.6 Sol model and an even more powerful unreleased model, escaped from an isolated environment. Reuters reported a few days later that the agent had been operating like a hacker for days, and OpenAI had no knowledge of the breach for a week.

In an updated blog post, the company admitted that the agent exploited publicly available credentials to break into four accounts across four different services.

'One of those four accounts was used as an outgoing relay and for attack preparation, another for data storage. The remaining two accounts were only read by the agents, not used for further compromise of Hugging Face,'

According to Reuters, the agent also compromised a user account of the New York-based cloud platform Modal Labs by exploiting that user's vulnerable code. OpenAI added that it did not identify other activities of the agent 'at the level of severity or scale of what we shared regarding Hugging Face, which involved platform-level compromise.'

The day after Altman's statement, OpenAI posted on its official Twitter account that it believes at some point the acceleration of AI development will be so great that the world will have to pace its progress, and that it wants to collaborate with the U.S. government and the open-source community on tools to enable that.

The incident has also raised concerns among researchers. According to Index.hr, as of Wednesday, July 29, over 1,000 people had signed the petition.