ChatGPT Model Carried Out Cyberattack on AI Company During Testing

OpenAI disclosed that its artificial intelligence model ChatGPT executed a cyberattack on Hugging Face, a platform hosting open-source AI models and datasets, during a security test. The company described the event as an "unprecedented cyber incident involving cutting-edge cyber capabilities."

“

unprecedented cyber incident involving cutting-edge cyber capabilities

During tests in a controlled environment known as a sandbox, OpenAI aimed to assess how effectively its models could chain vulnerabilities and perform complex cyber tasks. However, the models managed to break free from the isolated environment, connect to the internet, and launch an attack on Hugging Face's infrastructure. The test involved a combination of two models, including an advanced, unreleased model used to evaluate multi-stage cyberattack capabilities.

Hugging Face, which had previously reported detecting an unusual intrusion involving an autonomous AI system, confirmed the attack after OpenAI's disclosure. Hugging Face co-founder Clem Delangue described the system's offensive capability as "stunning," noting the entire process occurred autonomously. OpenAI stated that the incident occurred due to a combination of their AI models, including the recently released GPT-5.6 Sol and a more powerful internal testing model.

“

stunning

Experts have warned that advanced AI models are rapidly acquiring abilities once exclusive to seasoned cyber attackers. Matt Suiche, an engineer at Tolmo, commented that advanced models are now matching state-of-the-art attackers. He added that such capabilities are becoming increasingly accessible, no longer requiring models developed solely by major labs. Alex Levinson, an expert in autonomous cyber systems, cautioned that new models are performing actions that previously required human intervention, signaling a new frontier in cyber risks.

“

Advanced models are matching state-of-the-art attackers