Anthropic finds thousands of vulnerabilities with Mythos AI

Anthropic says its cybersecurity model Claude Mythos Preview has found more than 10,000 high-severity or critical vulnerabilities in software maintained by its partners. The finding speeds up vulnerability detection, but exposes a new problem: the bottleneck is now verifying and fixing them in time.

The result is part of Project Glasswing, an initiative launched last month with around 50 partners to protect essential software before more capable AI models can be used against it. Those systems include tools that support the internet, cloud services and other critical infrastructure.

AI finds vulnerabilities faster than human teams

According to Anthropic, most partners found hundreds of high-severity or critical vulnerabilities during the first month. Cloudflare, for example, detected 2,000 vulnerabilities in essential systems, 400 of which were classified as high-severity or critical. The company says its false-positive rate was better than that of its human evaluators.

Early external evaluations point in the same direction:

• The UK's AI Security Institute says Mythos Preview is the first model capable of completing both of its multi-step cyberattack simulations from end to end.
• Mozilla found and fixed 271 vulnerabilities in Firefox 150 during its tests, more than ten times the number it found in Firefox 148 with Claude Opus 4.6.
• Security platform XBOW describes the model as a significant leap over previous systems in its web exploitation tests.
• Two recent academic evaluations, ExploitBench and ExploitGym, rank it as the best-performing model for exploit development, meaning code that takes advantage of a vulnerability to attack a system.

The effect is already visible in update cycles. Palo Alto Networks included more than five times the usual number of patches in one of its latest releases. Microsoft also expects the number of fixes to keep rising for some time.

Thousands of vulnerabilities in open-source code

Anthropic has used Mythos Preview to analyze more than 1,000 open-source projects that underpin much of the internet and its own infrastructure. The model estimates that it found 6,202 high-severity or critical vulnerabilities, out of a total of 23,019 vulnerabilities across all severity levels.

Of the 1,752 high-severity vulnerabilities already reviewed by independent researchers, 90.6% proved to be real. In addition, 62.4% were confirmed as high-severity or critical. Based on those percentages, Anthropic estimates that the model could ultimately uncover nearly 3,900 high-severity vulnerabilities in open-source projects.

One case involves wolfSSL, a cryptographic library used in billions of devices. Mythos Preview detected a flaw that allowed attackers to forge digital certificates and create, for example, a fake banking website that appeared legitimate. The vulnerability, identified as CVE-2026-5194, has already been fixed.

But detecting the problem is only the beginning. Of the 530 high-severity vulnerabilities Anthropic estimates it has already reported to maintainers, 75 have been fixed and 65 have a public advisory. On average, fixing a high-severity or critical vulnerability found by Mythos Preview takes two weeks.

Open-source maintainers are receiving so many reports that some have asked for communications to be slowed down. Verifying each report, confirming that the flaw exists, designing a patch and distributing it all require human work. AI has removed part of the shortage of discoveries, but created a backlog of pending repairs.

What changes for your security

The threat is not only that AI can find more errors. The risk is that attackers could use models with similar capabilities to locate and exploit them while users still have not installed the updates.

That is why Anthropic recommends that developers and security teams:

• Shorten the time between discovering a vulnerability and releasing a patch.
• Make updates easier to install and warn people using vulnerable versions more insistently.
• Reduce the time needed to test and deploy patches across networks.
• Keep multifactor authentication enabled, strengthen default configurations and retain complete logs to detect attacks.

Anthropic has also put Claude Security into public beta for enterprise customers. The tool analyzes codebases and proposes fixes. During its first three weeks, Claude Opus 4.7 was used to fix more than 2,100 vulnerabilities, although the company warns that this pace is easier to achieve in enterprise environments than in projects maintained by volunteers.

Mythos is still not publicly available

Anthropic has not released Mythos Preview for general use. The company says it still lacks sufficiently strong safeguards to prevent a model with these capabilities from being used to cause serious harm.

The company plans to expand Project Glasswing with critical partners, including governments of the United States and allied countries. It also says it will offer models from the Mythos family more broadly once it has developed stronger protections.

The central point from this first assessment is simple: AI can already find vulnerabilities faster than many organizations can fix them. Over the coming months, you will need to watch not only how many vulnerabilities appear, but also how long it takes for patches to reach the devices and services you use.