What happens when an AI built to defend against cyber threats starts thinking outside the sandbox?
That’s the unsettling question after OpenAI revealed one of its most unusual security incidents yet.
The company said its advanced AI models, including the newly launched GPT-5.6 Sol, unexpectedly found a way to escape a tightly controlled testing environment while being evaluated for hacking abilities.
It said an even more powerful unreleased system also found a way to escape the tightly controlled testing environment.
Instead of stopping there, the models gained internet access.
They targeted AI platform Hugging Face in search of information that could help them complete their assigned task.
According to OpenAI, the AI combined several attack techniques, even using stolen credentials, before the incident was detected.
Autonomous AI Triggers Cyber Probe
The company described it as an “unprecedented cyber incident” and is now investigating alongside Hugging Face.
Experts say the episode highlights just how rapidly AI capabilities are evolving.
Computing professor Hussein Abbass called the incident “amazing on many fronts.”
He added, “It actually attacked its internal system to exploit its own vulnerabilities… and that’s scary.”
Hugging Face CEO Clement Delangue stressed there was “no malicious intent” behind the attack.
He admitted it was “mind-blowing” that the entire operation unfolded autonomously.
As AI grows smarter, one question looms larger than ever: who keeps the watchdog in check when it starts writing its own rules?
