Anthropic published a first update on Project Glasswing, an initiative created to protect critical software from AI models that are increasingly capable of finding and exploiting vulnerabilities. In just one month, Claude Mythos Preview and around 50 partners identified more than 10,000 high- or critical-severity flaws.
The most important figure isn’t just how many errors AI can discover. The real challenge now is how quickly teams can verify each finding, notify those responsible and distribute patches. Has cybersecurity just reached a new bottleneck?
Claude Mythos Preview speeds up vulnerability discovery
Project Glasswing brings together organizations that develop software essential to the internet and other critical infrastructure. Fixing flaws in these products can reduce risk for thousands of companies and potentially billions of users.
According to Anthropic, most partners found hundreds of high- or critical-severity vulnerabilities during the first month. Some teams say their speed at finding errors increased more than tenfold.
Cloudflare, for example, reported 2,000 flaws in critical systems, 400 of which were classified as high or critical severity. The company also said its false-positive rate was better than that of its human evaluators.
The company acknowledges that these results cannot yet be published with all the technical detail. The standard practice in the industry is to wait around 90 days after discovering a vulnerability, or roughly 45 days after a patch exists. The goal is to give users time to update their systems before revealing information that could help an attacker.
For now, Anthropic is sharing aggregated statistics and illustrative examples. Full details will arrive once the patches have been widely installed.
External evaluations
The first results from independent evaluators point in the same direction:
- The UK AI Security Institute reported that Mythos Preview was the first model to complete both of its end-to-end, multi-step cyberattack simulations.
- Mozilla found and fixed 271 vulnerabilities while testing Firefox 150, more than ten times the number it had detected in Firefox 148 with Claude Opus 4.6.
- XBOW, an independent security platform, described the model as a major improvement over existing models in its web exploitation benchmark.
- The academic benchmarks ExploitBench and ExploitGym ranked it as the best-performing model for exploit development among the systems evaluated.
Anthropic also notes that several vendors are releasing more fixes than usual. A recent Palo Alto Networks update included more than five times the normal number of patches. Microsoft expects the number of new fixes to continue growing for some time, while Oracle says it is now finding and fixing vulnerabilities several times faster in its products and cloud services.
Thousands of flaws in open-source projects
Anthropic used Mythos Preview to analyze more than 1,000 open-source projects that support a significant part of the internet and its own infrastructure.
The model estimated 6,202 high- or critical-severity vulnerabilities, out of a total of 23,019 findings that also include medium- and low-severity issues. So far, independent researchers have reviewed 1,752 of the most serious flaws.
Of those evaluated vulnerabilities, 90.6% were confirmed as real. In addition, 1,094, or 62.4%, were ultimately classified as high or critical severity. Based on these percentages, Anthropic estimates that Mythos Preview could end up discovering nearly 3,900 serious flaws in open-source software, even if it finds no new problems.
One notable case appeared in wolfSSL, a cryptography library used by billions of devices. Mythos Preview built an exploit that could allow digital certificates to be forged. In practice, an attacker could have created a fake website for a bank or email provider that appeared legitimate to the victim.
The vulnerability has already been fixed and received the identifier CVE-2026-5194. Anthropic says it will publish the full technical analysis in the coming weeks.
The problem is no longer finding errors, but fixing them
The review process requires reproducing each vulnerability, confirming its impact, checking whether a solution already exists and preparing a detailed report for maintainers. In open-source projects, this work often falls to small teams or volunteers.
Several maintainers told Anthropic that they are overwhelmed and need more time to design patches. On average, a high- or critical-severity vulnerability discovered by Mythos Preview takes about two weeks to fix.
So far, Anthropic estimates that it has reported 530 serious vulnerabilities to project owners. Of those, 75 have already been fixed and 65 have public advisories. There are also 827 other confirmed vulnerabilities that the company is trying to disclose as soon as possible.
The number of patches still appears low, but there are several reasons. Many findings remain within the 90-day coordinated disclosure period. In addition, some fixes are published without a formal advisory, meaning Anthropic has to detect them through additional analysis.
The difficulty reflects an important transition: AI can generate thousands of reports in a short time, but human teams still have to review, prioritize, fix and deploy each solution. If reports are inaccurate, the problem gets worse because maintainers lose time separating useful signals from noise.
What cybersecurity teams should do
Anthropic warns that models with capabilities similar to Mythos Preview will soon become more common. In the meantime, developers and network defenders can reduce risks with familiar measures that are now more urgent than ever:
- Shorten patch cycles and release fixes as soon as possible.
- Make updates easier so users can install them without complicated processes.
- Reduce the time needed to test and deploy patches within organizations.
- Harden default configurations for networks and systems.
- Require multifactor authentication for critical accounts and services.
- Maintain complete logs to detect incidents and respond quickly.
The recommendation is especially relevant because there is a dangerous window between the moment a vulnerability is discovered, the moment a patch is created and the date when most users install that fix.
Anthropic prepares tools for defense
Project Glasswing is not limited to the most capable models. Anthropic is also advancing tools based on models available to customers and security professionals.
The company launched Claude Security in public beta for Claude Enterprise customers. The tool analyzes codebases, identifies vulnerabilities and suggests fixes. During its first three weeks, Claude Opus 4.7 reportedly helped fix more than 2,100 flaws.
Anthropic also created a Cyber Verification Program so professionals who conduct vulnerability research, penetration testing and red-team exercises can use its models with certain safeguards adjusted for legitimate cybersecurity contexts.
For eligible customers, the company plans to offer the internal tools used with Mythos Preview. The package includes specialized instructions for repetitive tasks, a system that divides analysis among multiple agents, classification and report-generation features, as well as a threat-model builder.
Cisco, another project partner, published its Foundry Security Spec, designed to help other organizations create similar security evaluation systems.
The future of Project Glasswing
Anthropic has not yet publicly released Mythos Preview-class models. The company says it still does not have sufficiently strong safeguards to prevent misuse that could cause serious harm.
The concern is easy to understand: if AI can discover and exploit vulnerabilities at a fraction of today’s cost and time, attackers could take advantage of it too. That is why Project Glasswing aims to give defenders of essential systems an advantage before these capabilities become widely accessible.
Anthropic plans to expand the initiative with critical partners, including governments in the United States and allied countries. It also expects to offer Mythos-class models through a general release once stronger protections are in place.
The promise is significant: more resilient software, earlier error detection and fewer opportunities for attackers. But the transition period requires discipline. AI can already find flaws at extraordinary speed; now the industry must learn to process and fix them with the same urgency.