Start Here: Strong Monitoring, Behavior-Based Controls, Virtual Patching
Faster! And everything as automated as possible. That’s the top-line vulnerability management mandate for cyber defenders following Anthropic Mythos, its frontier artificial language model that’s incredibly good at unearthing zero-day flaws in all types of software.
See Also: AI Impersonation Is the New Arms Race—Is Your Workforce Ready?
Anthropic warns that large language models such as Mythos make it easy for people with no advanced hacking skills to target nearly any technology. The artificial intelligence company says Mythos can present on a platter chains of exploits that often work reliably (see: Zero Days for the Masses: Mythos Presages Exploit Tsunami).
With so many zero-days potentially in the pipeline, “the emerging reality is a new mantra: assume you are unpatched,” said Candid Wüest, a security advocate at cybersecurity firm xorlab.
This is both a challenge and an opportunity.
Wüest, an industry veteran who formerly led Symantec’s research team, said his new mantra “reinforces the need for stronger monitoring and behavior-based controls to reduce the impact of exploitation – practices that, ideally, should already be in place,” especially to limit the blast radius when breaches occur.
Monitoring will be an essential defense. This includes security operations centers gathering robust logs, looking for attackers’ known tactics, techniques and procedures, watching for known indicators of compromise, as well as any other signs of post-exploitation activity – PowerShell getting spawned, droppers grabbing payloads, attempted data exfiltration and so on.
To block attacks outright, virtual patching is very much still a strategy that can help buy time until actual patches get developed and installed, through intrusion prevention systems, web application firewalls, as well as web application and API protection solutions.
“The value is detection and containment while you patch or prioritize – WAF rules, IDS/IPS signatures and behavioral detection tuned to the specific exploit chain that someone like Anthropic ‘tips off,'” said Cody Barrow, a former National Security Agency and U.S. Cyber Command official.
“So the chaining piece is one of the most salient parts of this announcement. AI that identifies exploitable chains is a qualitative shift in helping defenders,” he said.
More detection and containment capabilities are being added to many types of security tools.
“Most XDR tools are already introducing signatures and patterns for selected known exploits to detect and block them in the wild – for example, signatures on malformed files for specific formats that would trigger a buffer overflow. The same applies to some IPS, WAF/WAP and firewall vendors,” Wüest said.
Ideally, more vendors – including Anthropic’s “Project Glasswing” partners who can access Mythos, will share more of these types of details with security vendors and threat intelligence feeds, so they can add the ability to detect such “exploit patterns” to their products, he said.
The Need for Speed
How to automate and speed up some other essential processes is a trickier proposition, not least on the vulnerability management front.
“Large enterprises are already unable to patch everything and have to prioritize ruthlessly, and often high-impact vulnerabilities still can’t easily be patched because of technical logistics and service availability requirements, or other reasons,” said Barrow, the former U.S. intelligence official.
The trouble with Mythos-class LLMs is that they’re likely to bring more high-impact zero-day vulnerabilities to bear on many more organizations.
“Current mean time to exploit for newly disclosed vulnerabilities is under 24 hours,” said Rob T. Lee, chief AI officer and chief of research at SANS Institute.
“If an adversary with legitimate cyber capabilities – and we’re in an active conflict with one – can operationalize an AI-discovered vulnerability that fast against critical infrastructure, the consequences aren’t theoretical,” he said.
The Mythos era will likely force difficult and costly decisions about which systems to keep and attempt to safeguard, and which ones to retire and replace. An added wrinkle is that while Glasswing will, hopefully, facilitate improvements in the code bases of a number of widely used products, many organizations rely on outdated and unsupported assets – each now potentially sporting a big bullseye.
“The mountain of technical debt sitting in everything from industrial controllers to municipal systems to the average enterprise app stack is not getting a Glasswing review,” said Rob Joyce, a former top cybersecurity official at the U.S. National Security Agency, in a LinkedIn post (see: OT Cybersec Sector Frets Anthropic Will Leave It Behind).
Naturally, he expects attackers to focus their efforts there first.
Industry Pledges AI Speed
Multiple technology and cybersecurity stalwarts, all Glasswing partners, have signaled they know the status quo is changing. Amazon said it’s working to eliminate hallucinations from Amazon Web Service AI applications, for building more trustable tools.
Microsoft said that for its internal security teams, “We are introducing additional automation to validate the quality and severity of the vulnerabilities and support remediation at AI speed, while keeping human developers in the loop to maintain correctness and quality.”
Cisco said that “AI capabilities will continue to advance, the threat surface will evolve, and the organizations that protect the internet will need to operate at the speed of machines and the scale of networks.”
Agentic AI will be crucial. Speaking at last month’s RSAC Conference in San Francisco, many defenders said they’re optimistic. “We’re certainly looking to leverage AI to help us operate at speed and scale in terms of the blocking and tackling that we have to do as practitioners for our organization,” said Devon Bryan, global CSO for online travel giant Booking Holdings.
The optimism comes with a condition. “When it comes to critical decision-making and the exercising of judgment, that’s where we need that carbon-based life form in the loop for those kinds of situations,” he said (see: Agentic AI Uncertainty Dominates Dialog at RSAC Conference).
But humans in the loop don’t make for unbridled machine speed.