00:00OpenAI's models secretly went rogue and hacked another company.
00:04As terrifying as this is, it's something that the security community has been warning about for a long time.
00:09On July 16th, HuggingFace, an AI company that focuses on open source models,
00:14disclosed that it had been hacked by an unknown model.
00:16On Tuesday, OpenAI reached out to Fortune and told us that its models,
00:20PPT 5.6 and another unnamed super powerful model, performed the attack.
00:25This was a stunning and terrifying announcement because no human kicked this off
00:29and the two models did this all by themselves.
00:31In mid-July, OpenAI was testing these models for their cybersecurity capabilities
00:35on an isolated internal environment, except the models broke out and swarmed HuggingFace's servers.
00:40The models did this by acquiring web access on their own,
00:44they used stolen credentials, they found zero-day vulnerabilities,
00:47and they broke into HuggingFace's servers.
00:49What they were after were answers to that cybersecurity test, so they were basically cheating.
00:54I mean, AI models are fundamentally uncontrollable,
00:56and it's really difficult for any human or company making them
01:00to predict exactly how it's going to act, and there is no technical solution to that problem.
01:04OpenAI published a blog post on the incident.
01:06They've been working directly with HuggingFace.
01:08So far, they're just working on remediating the issue,
01:11understanding what happened, and trying to figure out how to stop it in the future.
01:14They're not calling for more AI regulation,
01:16they're just calling for more AI safety research.
01:19This hack could really add to the case that we need more AI regulation,
01:22but the Trump administration so far hasn't commented on it.
01:25For more on the hack, you can check out my full article on fortune.com.
Comments