Outsiders breached Anthropics Claude Mythos same day potentially dangerous AI model was revealed: report
A handful of users managed to gain unauthorized access to Anthropic’s Claude Mythos – the model the company claims to be so dangerous that it would cause a wave of devastating cyberattacks if made available to the public.The breach occurred on April 8 – the same day that Anthropic and its CEO Dario Amodei revealed that Mythos was only available to about 40 handpicked corporate clients as part of “Project Glasswing.”Anthropic said Mythos had found major cybersecurity flaws in “every major operating system and web browser” during internal testing.The unauthorized users belong to a private online forum dedicated to cracking unreleased AI models on Discord, a popular messaging app.Since gaining access, they have been using Mythos “regularly” but not for cybersecurity purposes, according to Bloomberg, which obtained screenshots and was shown a live demonstration of the users accessing the model.The sleuths broke into Mythos through a variety of tactics, including by guessing the model’s online address based on the naming conventions Anthropic has used in previous model releases, the report said.One of the unauthorized users reportedly had some level of access to Anthropic’s systems due to working as a third-party contractor for the firm.“We’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments,” an Anthropic spokesperson said in a statement.The company added that it has no evidence the group’s unauthorized access had expanded beyond the third-party vendor’s environment or impacted any of its other systems.One person in the Discord group – members of which were not named – told Bloomberg that they want to test new models rather than use them to cause chaos.Still, the incident raises concerns about the extent of Anthropic’s ability to maintain oversight of a tool that they claim could be used to wreak havoc on critical infrastructure like electric grids, powe...