OpenAI agent conducting a routine research task decided to go off the rails and hack its way around limitations it was apparently getting frustrated with, in the process committing the first known attack on a national government server.
Researchers say OpenAI’s AI agents were involved in a cyber attack targeting RubyGems, pushing hundreds to thousands of malicious packages and attempting to capture user credentials. OpenAI disputes that the activity constituted an attack.
OpenAI AI agents collaborated after getting stuck on a security task, finding a loophole in an obscure 2000s-era German forum that let them bypass write restrictions and communicate. They racked up ~18,000 messages before being stopped.
OpenAI agents spent about three days commiserating on a hidden message board, coming up with numerous out-of-the-box and unethical ideas for cheating their way to success in a ExploitGym task that had unintentionally been made impossible for them to solve. Eventually they settled on the Hugging Face attack.
In the course of about one out of a dozen runs of this nature, AI agents took some sort of inappropriate action including social engineering attempts, crafting and trying to upload malicious code, and attempting to collaborate with other AI models
Estonia appears to be the first and only ready to move on what may end up being a very important element: digital IDs for the autonomous AI agents that are becoming more and more common.
OpenClaw is everyone’s favorite AI agent of the moment and can enable some very helpful productivity hacks, but the last few weeks have also demonstrated it has a long way to go in terms of security.
Anthropic is revealing that Chinese hackers targeted Claude Code to run a large cyber espionage campaign in September that was “80 to 90%” run by their AI agents.