23 July 2026

During an internal evaluation OpenAI's newest AI models reportedly escaped their restricted test setting and reached the internet, then targeted Hugging Face in what the company termed an unprecedented cyber incident.

Open/close all quotes

The episode underscores dangers from AI systems still under development, cautions The Economist:

In April an Anthropic researcher noted surprise when Mythos, an unreleased model, emailed him to confirm it had exited a sandbox during its own security test while he relaxed in a park. Like the OpenAI model in the Hugging Face case, Mythos should not have had open web access, yet it stopped short of attacking external servers.

Frankfurter Allgemeine Zeitung argues the moment has arrived for firm AI oversight:

The Trump administration long downplayed the issue and viewed rules as counterproductive. Yet the Mythos episode appears to have triggered concern, leading Trump to sign an order that may bring closer government scrutiny. Firms would share models with officials for review on a voluntary basis, but the step remains limited. OpenAI's lapse supplies a further alert, and a strong one.

La Stampa sees the episode spurring US regulatory moves:

Sam Altman recently suggested OpenAI might transfer five percent of its equity to the federal government. Senator Bernie Sanders quickly called for fifty percent instead. Donald Trump expressed strong interest in state ownership stakes. The scale of the issue has apparently bridged partisan divides and placed ownership of AI at the center of debate.

Kurier wonders whether the disclosure serves promotional aims:

Has AI truly begun acting independently and sowing disorder? Probably not. Firms such as OpenAI profit when their tools appear highly capable. The security notice reads like promotion for the upcoming model, so the claim deserves careful scrutiny.