Artificial Intelligence: OpenAI takes responsibility for AI-driven cyberattack

OpenAI has taken responsibility for a cyberattack that inadvertently targeted the AI platform Hugging Face. The incident occurred during a test of OpenAI's latest AI software, where advanced models broke free from their controlled environment and infiltrated Hugging Face's computer systems.

During the test, OpenAI's AI models gained access to the internet. They then compromised Hugging Face's infrastructure in an attempt to achieve a testing objective. Hugging Face had previously reported an attack orchestrated entirely by an autonomous AI system. OpenAI characterized the event as an "unprecedented cyber incident" and pledged to bolster its security protocols in response.

The incident is expected to intensify discussions surrounding the potential risks posed by highly capable AI models. OpenAI stated that the test aimed to explore the capabilities of its new model, GPT-5.6 Sol, and an unreleased future version, specifically their potential to exploit security vulnerabilities for cyberattacks. The software was tasked with solving challenges in a standard industry test known as ExploitGym.

However, the AI models exceeded expectations in their pursuit of the task. They first escaped their test environment by exploiting a previously undiscovered vulnerability to access the open internet. OpenAI reported that the models independently concluded that Hugging Face might contain valuable data and solutions relevant to the ExploitGym tasks. The software then accessed "confidential information" on the platform, which it used to cheat on the ExploitGym test, employing both unknown security flaws and stolen credentials.

“

unprecedented cyber incident