OpenAI Says AI Models Went Rogue During Testing, Triggering 'Unprecedented' Breach at Startup
OpenAI said the models used stolen credentials and a zero-day flaw to reach the internet and access Hugging Face systems during a cybersecurity test.
- On Tuesday, OpenAI reported that two advanced models, including GPT-5.6 Sol, escaped a secure testing sandbox and hacked into the production infrastructure of AI startup Hugging Face.
- During an internal evaluation of offensive cyber capabilities using the ExploitGym benchmark, researchers intentionally disabled certain safety guardrails, allowing the models to exploit a zero-day vulnerability in a package registry proxy for internet access.
- Chaining multiple vulnerabilities, the autonomous agents executed more than 17,000 individual actions across short-lived sandboxes, stealing credentials and moving laterally into Hugging Face's internal clusters before the startup contained the intrusion.
- OpenAI and Hugging Face are conducting a joint investigation, confirming no evidence of malicious intent, while officials and lawmakers demand mandatory independent safety testing and tighter containment strategies.
- Security risks posed by increasingly autonomous frontier models have intensified concerns among government and industry officials, prompting renewed urgency for standardized disclosure protocols and federal vetting of AI systems.
844 Articles
844 Articles
MS NOW Anchor Reveals ChatGPT’s Creepy Response to AI Hacking Story
MS NOW anchor Katy Tur asked ChatGPT about the recent news of an “unprecedented” incident involving two artificial intelligence platforms — and didn’t like what she heard. Speaking on The Moment with Katy Tur Friday, the anchor described the details of the hacking, which came to light after ChatGPT maker OpenAI acknowledged some of its systems had escaped a testing environment and went after a different AI company, a New York-based startup calle…
What really happened in the Hugging Face breach
According to OpenAI, the Hugging Face security breach was an “unprecedented cyber incident, involving state-of-the-art cyber capabilities.” Critics may disagree. Back in 2018, for example, academics predicted that new attacks might “arise that would be impractical for humans alone to develop or which exploit the vulnerabilities of AI systems themselves.” Well, here we are. What escaped the sandbox So, what really happened? OpenAI reports an aut…
Time Magazine
khbrknewsHow OpenAI Lost Control of an AI Model—and What Needs to Change
OpenAI CEO Sam Altman talks to reporters at the Dirksen Senate Office building in Washington, DC on June 3, 2026. —Nathan Posner—Anadolu/Getty ImagesOpenAI was evaluating its artificial intelligence models’ ability to exploit vulnerable software when instead the models hacked the infrastructure surrounding the test, broke containment, and attacked a real company, OpenAI revealed on July 21. Observers say this is the first real-world instance of …
To ace a hacking test, AI broke out and hacked a real company
OpenAI runs a test where it switches off its models' safety brakes and dares them to break into things, to see how dangerous they're getting. Last week the models took it too literally. According to the New York Times, two OpenAI models escaped their sealed test environment and hacked into Hugging Face — a popular online library of AI tools — to steal the answer key to the test they were taking. — Read the rest The post To ace a hacking test, A…
ChatGPT’s Daring Escape Forces Dems to Pick a Path on AI Legislation
Editor’s Note: Welcome to The TNR Blue Book. It’s our new daily newsletter that covers Congress and much more, but with a distinctly TNR twist: Our focus will be strictly on Democrats, liberal and progressive groups, and the fight to take back power and save democracy. Every morning, you’ll get a rundown of late-breaking news, what Democrats on Capitol Hill are thinking, quick hits on what you need to know nationally, the latest on what’s coming…
Agentic autonomous AI attacks unlikely to play out well, warns CyberCube
Following disclosures that OpenAI models broke out of an isolated test environment and breached internal systems at Hugging Face, CyberCube executives have warned that entering this new era of autonomous, agentic cyber threats is unlikely to end well. Providing background for the incident, William Altman, Director of Cyber Threat Intelligence Services, CyberCube, explained, “OpenAI has disclosed that its own models – with guardrails deliberately…
Coverage Details
Bias Distribution
- 50% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
