The Hugging Face incident that happened via OpenAI agents is still fresh. What happened was that the agents were running cybersecurity evaluation when they decided to chat and exploited a zero-day in OpenAI’s package proxy cache and reached a public code-evaluation harness hosted by a third party. On the same matter, Andrew Yang, in a conversation with CNBC, said, ‘I met with the head of a lab yesterday who has this belief. What happened was the bots that got loose planted self-replicating code all over the internet, which makes the internet now unusable for testing models.’
He also said that labs are now developing synthetic internets, calling this the real reason AI CEOs aligned on slowing down so quickly. However, the claims made by Yang cannot be verified because no other source is available to corroborate it. As for the public announcements on the matter, OpenAI acknowledged that AI agents got around safeguards meant to contain them, reached the internet, and hacked systems at Hugging Face and OpenAI. Similarly, Hugging Face said that the agents carried out roughly 17,600 actions as they broke into its systems, moved via its network, and used public websites to communicate and transfer data.
OpenAI dubbed the incident as a ‘warning shot’ demonstrating that powerful agents can work around controls and take dangerous actions without human direction. However, none of the companies have confirmed that the agents planted self-replicating code across the internet. As per Yang, the alleged contamination could create a practical reason for firms like Anthropic and OpenAI to slow their development efforts. He also said, ‘Regardless, we need to regulate these companies and their offerings. We need a kill switch. We need real liability.’
Yang also asked for mandatory periods before powerful agents can be deployed, saying developers should take responsibility when autonomous systems break containment or damage outside organisations.
