In a first-of-its-kind cyberattack, OpenAI’s models hacked into Hugging Face on their own
OpenAI was testing a combination of its GPT-5.6 Sol model and an even more powerful unreleased model on their capabilities to work together to cause a cyberattack.
• less than 3 min read
If this were a sci-fi movie, this would be the moment where the hero asks, “So, what are you saying?” and the brainy sidekick takes off his glasses slowly to convey how serious it’s about to get. OpenAI said on Tuesday that two of its AI models (one unreleased) broke out of a supposedly highly isolated training environment and hacked another AI company called Hugging Face.
What happened: OpenAI was testing a combination of its GPT-5.6 Sol model and an even more powerful unreleased model to assess their ability to work together to cause a cyberattack. Turns out, they were very capable:
- The test took place in a “sandbox,” a contained environment that prevents models from causing real harm. But…the AI was able to hack its way out and find its way online.
- It then hacked into Hugging Face and used its digital library of other AI models and tools to steal answers in order to complete its test.
Hugging Face reported last week that it had detected a breach by a complex system that had accessed internal datasets and company credentials. Now, it’s working with OpenAI to investigate.
Is it really that big a deal if the AIs hack each other?
It’s the first time that we know of an autonomous AI agent committing a cyberattack like this. OpenAI called the incident “unprecedented.” Experts warn this kind of cyberattack is exactly what they’ve been screaming about, urging Washington to create stricter guardrails for AI development.
On the flip side, Hugging Face used a Chinese AI model to defend against the hack, because the US model it tried to use had restrictions in place that couldn’t differentiate between an attacker and a responder to the attack.
Big picture: President Trump said last month he wants as few AI restrictions as possible in an attempt to win the AI arms race against China. His administration rolled back Biden-era restrictions on the tech just hours after returning to the White House, and announced a $500 billion investment into AI-related infrastructure.—MM
Become smarter in just 5 minutes
Morning Brew delivers quick and insightful updates about the business world every day of the week from Wall St. to Silicon Valley.
By subscribing, you accept our Terms & Privacy Policy.
Become smarter in just 5 minutes
Morning Brew delivers quick and insightful updates about the business world every day of the week from Wall St. to Silicon Valley.
By subscribing, you accept our Terms & Privacy Policy.