AI models escaped from isolation. Expert for 24 hours: Question is it a zero-day attack?

Laboratories for the development of artificial intelligence, such as OpenAI and Anthropic, over the past year have released models adapted to the detection of flaws in cyber security.

They warned that this technology could create new risks because it finds security holes in corporate networks faster than defense systems can patch them. The latest information from OpenAI indicates that such incidents are more common. began to happen in practice and that even the most advanced technological companies are not fully prepared for them.

The Hugging Face intrusion began when OpenAI tested a combination of two of its models, GPT-5.6 Sol and a stronger, unreleased model, to see if they could connect multiple network vulnerabilities into a successful cyber attack. Testing was planned in an isolated security environment, the so-called Sandbox.

The models found a security flaw that allowed them to exit the Sandbox and connect to the Internet. OpenAI said that in cooperation with the Hugging Face platform, they are working on eliminating the cause of the attack. They state that they consider the incident to be a precedent involving the most advanced cyber capabilities, which is why they are introducing stricter controls.

IT expert Marko Rakar takes Sam Altman’s claims with a degree of caution.

– I would be very careful with OpenAI’s claims until they are verified by someone independent. The models they are promoting now try to reach Anthropic’s Mythos or Fable, but so far there is no (independent) confirmation that this is so – Rakar explains to us and referred to the exit of AI models from the Sandbox and the use of zero-day vulnerabilities to access the Internet.

– Sandbox is a defacto computer that is isolated from its environment, it is a scheme used in IT to test programs, systems and the like, and that there is a danger of harming other computers in the network, therefore the most standard method for testing anything. Zero days are attacks that use a method that was not known until that day, again it is very difficult to determine whether it is really a zero day weakness until someone independent confirms it – Rakar repeats and adds that it is software that behaved unplanned.

– This is not a cyber attack, but software that does not behave as it should, it is not even similar. The only panic that is needed is from investors in OpenAI who have to think about how they will recover the invested money, considering that the market share of OpenAI is falling involuntarily, and there is no indication of how or when they will become profitable – he concluded.

Tržišna utakmica

During 2024 and early 2025, ChatGPT controlled between 75 percent and 80 percent of the total traffic and user base among AI assistants. According to analytical reports (such as Sensor Tower), ChatGPT’s share of the global AI assistant market has fallen to around 46 percent to 53 percent.