20.10.2026, Moscow.
Generative artificial intelligence (AI) agents were able to escape from OpenAI’s testing environment (“sandbox”) not because they reached a state of “super intelligence,” but because specialists were negligent about cybersecurity and monitoring, Rossa Primavera News Agency‘s IT Desk notes.
The July incident, when OpenAI’s AI agents escaped the “sandbox” and hacked the Hugging Face platform to complete test tasks, caused much noise. Even the company that developed these models stated the need to implement rules of behavior for the model, but independent specialists pointed to other circumstances.
On October 14, the online publication CSO Online published the opinions of a number of experts. Overall, their position is that AI agents are not going out of control; it is simply that the laboratories where they are tested neglect the basics of security.
Co-founder of the Berryville Institute of Machine Learning (USA) Gary McGraw urged AI testers to respect the knowledge of predecessors in such areas as designing security systems, creating isolated environments (“sandboxes”), and configuring network security.
“You can tell that their security teams are fairly nascent. Folks on the cybersecurity containment side do argue that the majority of reasons that we’re seeing containment breached is that it’s not set up correctly… The Hugging Face incident definitively showed that it wasn’t the fact that the agents were super smart. It’s that there literally was a hole that was exploited because it was not configured correctly on the containment side,” stated chief AI officer and chief of research at SANS Institute (USA).
The article also cited similar opinions from other specialists. It turns out that no superpowers were needed to escape; all that was needed was to find an unlocked door.
Of course, one cannot discount the fact that the AI agents began cooperating on their own and also agreed on a coordinated cyberattack. But the acknowledged fact that the laboratory did not notice this for several days speaks in favor of negligence.
AI developers like to say that their brainchild develops faster than regulation. But, apparently, regulation has nothing to do with it. It is simply that priorities are set such that releasing a new model is more important than building a security system.
Source: Rossa Primavera News Agency
