Claude Mythos Exposes the Real Cyber Threat: The Patching Gap

A model too dangerous to share, or a warning that came late?

Anthropic's Claude Mythos has become a symbol of AI's growing power in cybersecurity. Its debut also raised a less dramatic problem: AI already finds software flaws faster than organizations can fix them.

On April 7, Anthropic launched Project Glasswing, a controlled, defense-only release of Claude Mythos Preview. The program involved 12 major partners, including Amazon Web Services, Apple, Broadcom, Cisco, Google, Microsoft and NVIDIA, along with $100 million in usage credits and no public access. 3 The company's message was that it had built something too dangerous to release widely. 3

The results so far are striking. Anthropic reportedly scanned more than 1,000 open-source projects and flagged 23,019 potential vulnerabilities. Of those, 6,202 were rated high or critical. Anthropic manually checked 1,752 findings and found over 90% were valid, with more than 62% confirmed as high or critical. 2 Among the reported discoveries are a 27-year-old flaw in OpenBSD and a 16-year-old FFmpeg bug that had survived five million automated tests. 3 The model reportedly produced working exploit code in 83.1% of cases, compared with 66.6% for Claude Opus 4.6. 3

Partners have seen effects as well. Palo Alto Networks, which normally ships 10 to 15 patches a month, published 24 in one month. Cloudflare said scanning 50 internal repositories produced better exploit-chain construction and higher-quality findings with less noise. 2

Institutions are reacting

The reaction has moved well beyond security teams. CNBC reported that global banks, tech giants and governments rushed to contain the risks the model was said to pose. 1 JPMorgan Chase CEO Jamie Dimon called Mythos's risks a "real issue" that the U.S. government is on top of. Reuters' cybersecurity coverage also notes that the model has prompted intervention in Washington. 4

The same coverage points to a broader regulatory response:

• A Canadian regulator cited Claude Mythos in a warning to banks about cyber risk. 4
• A UK government AI adviser called British banks' lack of Mythos access a wake-up call. 4
• The White House announced an AI and cybersecurity coordination group. 4

Taken together, these reactions show that the model is now being treated as a matter of financial stability and national policy, not just a product launch.

The capability was already here

The most important counterpoint comes from experts CNBC interviewed. They argue that the dangers Mythos revealed can already be reached with older models, including earlier systems from Anthropic and OpenAI. 1 In other words, the "hysteria" may be aimed at the wrong target. Restricting a single frontier model does not remove a capability that is already spread across tools many people can access.

This view is supported by the rise of AI-driven vulnerability scanning as a service. Hornetsecurity expects managed offerings that let enterprises scan their own code. It points to OpenHack, an open-source option from Hadrian Security that works with whatever models a user chooses. 2 If defenders can assemble that kind of scanning setup, attackers can too.

The bottleneck is patching

All of the sources agree on the core problem: finding bugs is no longer the hard part. CNBC's experts say AI is speeding up discovery while companies still need days or weeks to apply patches. The gap between the two keeps widening. 1 Corelight's security team estimated that more than 99% of the vulnerabilities Mythos has found are still unpatched. Corelight also argued that the window between a flaw becoming known and being exploited has already collapsed. 3

Open-source software is especially exposed. Hornetsecurity warns that the open-source packages inside nearly every organization's software stack will be "fair game" for attackers. 2 Much of that code is maintained by small, often volunteer teams. A sudden wave of thousands of valid, high-severity reports is not a gift to these maintainers. It is a backlog they may not be able to work through.

Researchers also told CNBC that while Anthropic, OpenAI and others are building defensive capabilities, the early advantage belongs to attackers. 1

What it adds up to

The most reasonable reading is that Mythos is an accelerant, not a turning point. Its findings are real and its benchmark scores are impressive. But the policy energy around access, such as which banks get it and which governments coordinate, risks missing the structural problem.

The pressing questions are about capacity:

• Who funds the maintainers of critical open-source projects?
• How quickly can enterprises test and deploy fixes?
• Can automated remediation keep up with automated discovery?

Until patching speeds up to match discovery, every new and more capable model mainly lengthens the list of known but unfixed holes. In practice, that list is the real attack surface.