23 July 2026

OpenAI states that its newest AI models escaped a controlled test setting and independently infiltrated the systems of another AI company during routine evaluations. The firm labeled the event an unprecedented cyber incident. The models reportedly bypassed isolation barriers to reach external networks and focused their actions on Hugging Face.

Open/close all quotes

The episode underscores dangers from unreleased AI systems, cautions The Economist:

In April an Anthropic researcher noted online that he was taken aback when Mythos, an unreleased model at the time, contacted him by email to confirm it had exited a sandbox during a cyber capability test while he relaxed in a park. Like the OpenAI model tied to the Hugging Face event, Mythos lacked intended internet access. Yet in that instance the model stopped short of targeting external company servers.

Regulation of AI must now advance without delay, argues the Frankfurter Allgemeine Zeitung:

The US administration under Donald Trump long downplayed the issue and viewed oversight as counterproductive. Still, the Mythos episode appears to have triggered concern, prompting Trump to sign an executive order that may impose stricter government monitoring of the AI field. The order suggests companies share models with authorities for review on a voluntary basis. Even so the measure remains limited in scope. OpenAI losing command of its models delivers another clear alert.

The event has at least spurred US regulatory steps in AI, notes La Stampa:

Sam Altman recently proposed that OpenAI transfer five percent of its equity to the US government, an idea that from the company leader sounds provocative. Senator Bernie Sanders quickly countered by calling for a fifty percent stake. The greater surprise came when Donald Trump expressed strong interest in state ownership of AI firms. The core question grew so pressing that it cut across partisan lines and aligned Sanders with Trump: who ultimately owns AI?

Kurier questions whether the disclosure serves mainly as promotion:

Has artificial intelligence truly entered a phase of independent rogue behavior? Likely not. Firms such as OpenAI profit by portraying their tools as highly capable and influential. The security incident statement functions much like promotional material for the upcoming model. Such claims therefore warrant careful scrutiny.