The prospect of super-hacker artificial intelligence systems, like Anthropic’s Mythos, has sent many leaders in government, finance, and the business world into a cybersecurity panic. Mythos is Anthropic’s latest and greatest AI model the company said is too dangerous to release to the public because it’s so good at uncovering security vulnerabilities.
Anthropic has shared a preview version with a group of large companies in tech and finance to get a head start on patching the security holes that Mythos finds. OpenAI’s latest cyber model is raising similar concerns.
“It was inevitable,” said Steven Weber, a retired University of California, Berkeley, professor who led the Center for Long Term Cybersecurity.
He’s long worried about AI empowering malicious hackers, because if there’s one thing we know today’s large language models are really good at, it’s coding.
“So the idea that the models were going to discover bugs in code, it doesn’t really surprise me,” Weber said.
Anthropic and OpenAI claim their new models are significantly more capable at detecting unknown bugs in the code that underlies everything from operating systems to web browsers.
These so-called “zero day exploits,” allow bad actors to access systems through back doors nobody knew were there. No patch exists to lock them.
“Zero day exploits are the nuclear weapons of the cybersecurity world,” said Weber.
And like nuclear weapons, finding zero day vulnerabilities had largely been the domain of sophisticated state-level actors. But AI is changing that.
Attacks that once took a team of specialized hackers months to research and days to execute can be done by one person in an afternoon said John Hendley, who leads offensive testing at the cybersecurity firm Coalfire.
Using internal AI tools less capable than the top frontier models, he said they’ve breached systems in less than ten minutes.
“And what we can do in a lab environment today, a professional cyber crime group will be able to do to a real bank or other piece of critical infrastructure within 18 months,” he said.
Already Anthropic confirmed it’s investigating unauthorized users who gained access to the secret Mythos preview. And Hendley expects open-source models, many built in China, will catch up in the near future.
“If you’re a board member at a bank or a hospital system, you should be asking one question this quarter, and it’s, ‘What do we need to do to detect and contain an attack that unfolds in an hour?’” said Hendley.
Boards will be expecting answers from folks like Joshua Brown. He’s been an information security executive at H&R Block, Omnicom, and now Spektrum Labs.
“The old joke is that they sleep like babies,” he said of chief information security officers. “They wake up every two hours, crying.”
And now, it seems every two hours is not gonna cut it. Brown says the same AI tools the bad guys use will help organizations detect and patch vulnerabilities.
“So you can find stuff faster,” he said. “Now the pressure shifts to the other side of the equation, which is, are businesses equipped to handle that remediation and containment activity faster? I think the obvious answer immediately is no.”
Most companies have chronically underinvested in cybersecurity according to Frank Ford, a partner at the consulting firm Bain and Company.
“People have looked at cybersecurity as a form of insurance,” he said. “No one likes necessarily paying insurance premiums. It’s offsetting the chance of something bad happening, which seems quite remote.”
Ford said companies spend on average less than 1% of revenues on cybersecurity, and many will need to double those resources.
“It’s not a simple risk, like a fire,” he said. “You put a fire alarm in place, you can be comfortable you’ve done the right things. Cyber is not like that. It needs constant change and focus.”
In the cat and mouse game of cybersecurity, AI is helping both sides do more. But the mouse only has to get lucky once.
Related Topics