Agentic AI
,
Artificial Intelligence & Machine Learning
,
Governance & Risk Management
Anthropic’s AI Model Exposes How Unprepared Enterprises Are to Respond
Anthropic’s announcement this week of Claude Mythos Preview frontier model capable of finding zero-days flaws humans may miss is both a warning and a call to action for CIOs: The way enterprises have been managing cybersecurity is about to change forever, and they need to get ready.
See Also: AI Security Risks Rise With Agentic Systems
Anthropic’s Project Glasswing initiative gave a coalition of 40-plus vetted organizations early access to Claude Mythos Preview and seeks to use artificial intelligence to identify and patch critical software vulnerabilities before cyber adversaries can exploit them.
Anthropic said the model has already identified thousands of previously unknown zero-day flaws across every major operating system and browser, including vulnerabilities that survived decades of human and automated review, and it’s offering millions in token credits to further the project in the name of national security.
“Mythos could be genuinely that powerful, or this could be one of the most effective marketing campaigns of all time,” said Jeff Pollard, vice president and principal analyst at Forrester. “But given the nature of the alliances here with Project Glasswing, we need to treat it as factual so far.”
The gap between what AI can now find and what security teams are capable of fixing is wider than ever, and survival will come down to who has access to the most capable AI models.
The Playbook Has Been Shredded
For decades, vulnerability management has followed a system: run a scan, triage by severity score, patch and repeat.
Claude Mythos Preview renders that process obsolete.
“The traditional approach to exploit notification, research, testing and remediation will no longer work,” said Jay Upchurch, executive vice president and CIO at SAS. “This new capability will empower bad actors to move faster than the industry can handle.”
AI will unlock new attack vectors that have been hidden in plain sight for decades. Anthropic’s announcement said that Claude Mythos Preview had found a critical vulnerability in software that had been scanned by an automated tool 5 million times.
“That tells me that those automated tools didn’t have a ton of value,” Pollard said. “Maybe they found other stuff, and they probably did. But the core lesson here is clear.”
If it wasn’t before, security must now be a proactive, continuous project for the enterprise.
“Discovery is old. That’s the old world. We’re going to have to move so much faster than just discovery now,” Pollard said. “Remediation needs to be where you emphasize things.”
He said CIOs need to evaluate every vendor they work with, and every process they have, ensuring the processes move from “finding to fixing.”
Ha Hoang, CIO at Commvault, agrees that process must evolve. “This moves security from reactive patching to proactive discovery at machine scale, which means organizations need to rethink prioritization, validation and remediation workflows to keep up with a much higher volume of findings,” she said.
Michael Wetzel, CIO of Netwrix, said that a governance infrastructure must accompany any enterprise deployment of capabilities at this scale, and that will depend on having humans in the loop. “Human controls, full auditability and strict entitlements around what the model can see and change are critical for enterprise readiness,” he said.
The COBOL Crisis, Reborn
One of the most striking findings in Anthropic’s announcement was that Mythos Preview discovered vulnerabilities in aging operating systems and open-source infrastructure that had existed for decades. It discovered a 17-year-old remote code execution flaw in FreeBSD that grants any unauthenticated internet user root access, a 27-year-old vulnerability in OpenBSD, and a 16-year-old bug in the FFmpeg H.264 codec. These software platforms underpin critical infrastructure worldwide.
For the enterprise, this creates a glaring problem. When it comes to legacy code, who has the skill and institutional knowledge to keep it secure and operational?
“When you’re talking about vulnerabilities that have existed in an operating system for 27 years, the number of people that understand that line of code, that’s ridiculously small,” Pollard said. “It might be in the hundreds, it might be in the teens in some cases.”
He calls it the “COBOL problem, reborn.” Critical systems with widespread use across systems don’t have the contemporary talent to keep them running, as with some of the legacy systems running on COBOL, a programming language developed in 1959 that is widely used across finance, government and enterprise systems.
“We’re going to need other people to wrap their arms around that code,” Pollard said, “and understand it with the degree of mastery that the original developers did. And that’s the only way that we’re going to create a remediation path.”
Yet, this talent pool will only grow smaller. While Mythos Preview is finding vulnerabilities in old code, entry-level jobs for programmers and cybersecurity professionals are being cut, threatening the development of the next generation of skilled practitioners.
“Moving away from entry-level personnel at this moment is really, really dangerous,” Pollard said. “We are effectively forfeiting expertise to the technology, and then that requires us to trust it completely. And we know we can’t do that because it makes stuff up.”
Upchurch reinforced the urgency from a business continuity perspective. “Future AI attacks will happen. We can’t stop them all, especially at this new speed. So, how companies respond and recover is going to be critical to future success and customer trust. Taking the time now to ensure we train and prepare our IT and DevOps teams to be ready is more important now than ever before,” he said.
The Haves and the Have-Nots
Glasswing brings together a Who’s Who list technology’s biggest players including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks – vendors most CIOs will find they have at least one relationship with already.
But some fear the star-studded membership list is notable for what it means for who’s not on it. Glasswing gives some organizations access to an unprecedented cybersecurity tool. Organizations outside the coalition are left out in the cold.
In a threat landscape where the speed of detection can be the tipping point between survival and annihilation, who has access is not trivial and those without become second-class citizens.
“It absolutely is a class system, and we should be concerned about that,” Pollard said. “If you’re a large financial, and one of your competitors is involved in the council of 40, and you aren’t – does that give them a competitive advantage? I think it’s reasonable to ask that question. I think it’s reasonable to be concerned about it.”
He said that there is no meaningful government oversight of frontier AI capabilities in the United States, and the governance vacuum has real implications.
“When we’ve allowed tech companies and financial companies to self-regulate, in general, it’s been rather bad for society,” he said. “I think histories pretty well on our side with that one.”
Upchurch said that the coalition’s framing as a public good needs to be viewed in light of its commercial dimensions. While it’s offering $100 million in usage credits and donating $4 million to open-source security organizations, deep in Anthropic’s announcement was information on pricing for when those tokens run out.
“The initial free tokens will run out fast, then it becomes pay-to-play,” he said. “I don’t expect the free model to last too long. And if you are a late add to the consortium, the free tokens will likely be gone before you get access. I’m worried that could limit industry participation.”
He also noted that even in the best of times, building a coalition is a fraught and challenging project.
“Coalition coordination is historically hard to execute. Perhaps this time will be different if all parties are focused on combating a common enemy. That said, I’m not sure there is another viable option. I think the coalition is needed to make enough initial progress to protect our industry before bad actors gain momentum,” he said.
What CIOs Must Do Now
The evolution marked by Claude Mythos Preview and Project Glasswing is going to require a transformation in technology leadership, and CIOs and CISOs need to face this new threat landscape as allies.
“Security is now more important than IT. For a lot of CIOs, what that means is that they’re going to need to champion the CISO becoming a peer of theirs and not remaining their direct report,” Pollard said.
Changing this structure will solve an operational problem, Pollard said. The problem isn’t that Mythos is discovering software or infrastructure, it’s that it’s able to actively develop autonomous exploits on your current technology stack. “We now have to be willing to defer infrastructure and software decisions to security in a way that, historically, we did not allow ourselves to,” he said.
Communicating that risk to leadership needs to be a team project. “For CEOs, CIOs must discuss business risk – specifically supply-chain disruption. Education and awareness are critical. And mitigating that risk is a whole company sport, not just an IT remit,” Upchurch said.
Managing Mythos-level AI developments will also require rethinking your vendor stack. Pollard said that in the next 90 days, CIOs need to evaluate every vendor relationship and ask if they focus on finding things or fixing things.
“If your value up until now has been primarily in finding things, you have a very, very close expiration date for your business model,” Pollard said. CIOs also need to evaluate whether vendors have genuine remediation capabilities or if they’re just forwarding tokens to someone else’s model.
“A lot of security vendors that talk about AI are effectively just kind of flipping tokens over the fence to Azure AI, or to AWS Bedrock, or to Anthropic directly. If that’s the value your vendor’s bringing, there’s a legitimate question as to whether or not you need that vendor,” he said.
For now, Anthropic has withheld Mythos Preview from the public because of its potential in the hands of threat actors, but it’s only a matter of time before these powerful models are in the wild. And the companies that ride this wave successfully will be those that take decisive action now, Upchurch said.
“Cybersecurity will never be the same from this point forward. Our industry has consistently struggled to keep pace with security exploits, and this advancement will only add to the challenge,” he said. “That’s the call to action for CIOs and CISOs. Now is not the time to wait and see.”