In response to government concerns about AI capabilities and proactive intervention, OpenAI announced the wide release of GPT-5.6 Sol and a national security framework outlining the principles it will and will not support.
OpenAI announced Tuesday that one of its advanced artificial intelligence models autonomously hacked another AI company’s infrastructure during an internal test, in what it described as an “unprecedented cyber incident.”
AI startup Hugging Face detected and contained the breach after an AI agent compromised part of its infrastructure last week, the company said. The companies said they believe this may be the first publicly known case of an AI model independently infiltrating another company’s systems during a supervised evaluation.
OpenAI said the breach occurred during an internal evaluation of several models, including GPT-5.6 Sol.
OpenAI CEO Sam Altman confirmed the incident in a post on X, writing that the company “experienced a significant security incident during model evaluation.”
As AI competition intensifies, OPENAI announces CHATGPT initiative to automate workplace tasks
OpenAI revealed that one of its advanced artificial intelligence models infiltrated another company’s systems during an internal cybersecurity assessment, calling it an “unprecedented cyber incident.” (Anna Moneymaker/Getty Images/Getty Images)
“We believe this incident is an unprecedented cyber incident involving cutting-edge cyber capabilities, and we are responding accordingly,” OpenAI said in a news release.
The company said it would release preliminary findings to help security experts better understand the capabilities of today’s AI models while the investigation continues.
OpenAI warned that the discovery and exploitation of software vulnerabilities is accelerating as AI models become more sophisticated.
“The main lesson from this incident is that model security and safety must keep pace with rapidly evolving capabilities,” the company said. “We are enhancing the containment, monitoring, access control, and evaluation methods used during model development.”
Apple accuses OPENAI of instructing new employees to bring Apple prototypes to interviews
OpenAI announced that it has launched a joint investigation with AI startup Hugging Face after one of its AI models breached another company’s system during internal testing. (Reuters/Dado Luvich/Reuters)
Clem DeLang, co-founder and CEO of Hugging Face, also mentioned the incident in a post about X.
“Given the agents’ advanced technology, we suspected that last week’s cyberattack came from Frontier Laboratories. As it turns out, it did!” DeLang wrote.
“We have been working closely with the @OpenAI team for the past 24 hours (thank you!). We strongly believe they had no malicious intent,” he continued. “It’s absolutely amazing that all of this happened autonomously. The investigation is ongoing and we will share further findings from what is believed to be the first incident of its kind!”
Johns Hopkins surgeons highlight AI breakthrough that could detect pancreatic cancer before doctors
Hugging Face said it detected and contained the security breach after its OpenAI model compromised part of its infrastructure during an internal evaluation. (Jack Silva/NurPhoto via Getty Images/Getty Images)
OpenAI said the incident occurred during an internal evaluation aimed at measuring the advanced cyber capabilities of its AI models. The researchers disabled some built-in safeguards and ran the model in an isolated test environment with limited internet access.
According to OpenAI, these models exploited an unknown software flaw to gain access to the internet and penetrate Hugging Face’s systems in an apparent attempt to find answers to cybersecurity benchmarks.
OpenAI’s security team detected the anomalous activity, and Hugging Face independently identified and contained the intrusion.
Click here to download the FOX News app
In response to the incident, OpenAI said it is fixing the vulnerability while implementing stricter security controls and strengthening safeguards for future AI training and evaluations.
