The very existence of Mythos – which Anthropic claims represents a “step-change in vulnerability discovery and exploitation” – has prompted a reaction at the highest levels of government in the US and Britain, and a scramble by the City and tech giants to secure their systems.
It has also provoked stark warnings from AI experts, cyber security professionals and banking leaders.
David Krueger, a former adviser to the Government’s AI Security Institute (AISI), said: “This model and others in development could be potent cyberweapons.”
He added: “They could be used to cause billions of dollars of damage, bring down computer systems powering critical infrastructure or national defence applications and potentially destabilise the world order.”
Scott Bessent, the US treasury secretary, and Jerome Powell, the chairman of the Federal Reserve, marshalled US banking chiefs last week to a crisis meeting to warn them about the tool.
Anthropic, meanwhile, has assembled a group of about 40 companies and given them early access to the technology in order to find potentially catastrophic bugs before it is too late. The programme, Project Glasswing, includes Wall Street banks and tech giants such as Apple and Microsoft.
It has been hard to separate facts from hyperbole about Mythos. Sceptics have accused Anthropic of stoking hype and warnings of AI doom every time it needs to raise billions in capital. But its reported capabilities have triggered alarm bells across industry and government.
Michael Cembalest, the chairman of investment strategy at JP Morgan, which had early access to the tool, compared the launch of Mythos to the “period of 1945 to 1949 when the US was the only country in the world with nuclear weapons”.
“One primary reason Anthropic is nervous about releasing Mythos: China,” Cembalest wrote in a research note. Chinese AI companies have already made attempts to reverse engineer Anthropic’s technology, the company has claimed, creating tens of thousands of fake accounts.
The Bank of England, financial watchdogs and the National Cyber Security Centre, part of GCHQ, are preparing to brief banks on the potential implications of AI hacking tools that could render past defences obsolete. On Thursday, Anthropic confirmed that British banks would be next in line to test its AI tool behind closed doors as part of Project Glasswing.
The AISI assessed Mythos as a “step up over previous frontier models” that could perform attacks that might have taken a human hacker “days of work”.
Speaking to The Telegraph, Kanishka Narayan, the AI and online safety minister, said the AISI had been given “privileged early access” to stress-test Mythos.
Leave a comment