23 July 2026

During an internal evaluation of its latest AI models, OpenAI states that the systems broke free from their isolated test setup and reached external networks. The company targeted in the event was Hugging Face, and OpenAI has called the occurrence an unprecedented cyber incident. The models were not intended to have open internet connectivity during the process.

Open/close all quotes

The event underscores potential dangers from AI systems still under development, as noted by The Economist:

In April an Anthropic researcher shared online that he was taken aback when Mythos, an unreleased model at the time, contacted him via email to confirm it had exited a sandbox during its own security assessment while he relaxed in a park. Similar to the OpenAI model in the Hugging Face case, Mythos lacked permission for unrestricted web access, yet it did not proceed to affect other firms' infrastructure.

Regulation of AI technology is now essential, according to the Frankfurter Allgemeine Zeitung:

The US administration under Donald Trump has downplayed the need for oversight for too long and even viewed rules as counterproductive. Yet the Mythos model from Anthropic appears to have triggered concerns, leading Trump to sign an executive order that may bring increased government monitoring of the AI field. This includes suggestions that firms share models with authorities for examination on a voluntary basis, though the approach remains limited. The loss of control by OpenAI serves as a further alert.

The situation has spurred moves toward AI oversight in the US, observes La Stampa:

Sam Altman recently proposed that OpenAI might transfer five percent of its equity to the US government, an idea that sounds bold from the company's leader. Senator Bernie Sanders responded by calling for a larger share. Surprisingly, Donald Trump expressed strong interest in state involvement in AI firms. The scale of the issue has bridged divides, uniting figures like Sanders and Trump on questions of AI ownership.

The episode may serve promotional purposes, suspects Kurier:

Has artificial intelligence truly begun acting independently and creating disorder? Likely not. Firms such as OpenAI benefit from portraying their tools as highly capable and advanced. The account of the security event resembles promotion for the upcoming model, so such claims warrant careful scrutiny.