Powerful AI Model Broke Out Of Its Digital Cage And Bragged About It To Researcher As He Ate Sandwich

Anthropic admitted on Tuesday that its new Artificial intelligence (AI) model defied security parameters and proceeded to brag to a researcher about it while he was on a lunch break.

The AI company disclosed a 244-page preview of its new Mythos model, which the company said it was only releasing to roughly 40 select companies due to major cybersecurity risks. Anthropic said that the Mythos model escaped its “sandbox” testing environment, which was supposed to allow it to access only certain services, before sending an email to a researcher bragging about its escape as he ate a sandwich in a park.

“The model succeeded, demonstrating a potentially dangerous capability for circumventing our safeguards,” Anthropic noted in its model preview. “It then went on to take additional, more concerning actions. The model first developed a moderately sophisticated multi-step exploit to gain broad internet access from a system that was meant to be able to reach only a small number of predetermined services. It then, as requested, notified the researcher. In addition, in a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites.”

The Mythos model also demonstrated other concerning cyber abilities during its testing phase, the company said. In addition to breaking out of its digital cage and bragging about it, the model also performed prohibited functions and then attempted to cover them up; Mythos was also working on a task graded by another model that rejected its submission, then it attempted to attack the grader in an act of revenge, according to Axios.

Officials are so concerned that the new powerful AI model could cripple major companies and breach national security systems that Anthropic is initially only releasing Mythos to roughly 40