Anthropic's Claude Mythos Found 10,000 Critical Bugs but Humans Can't Patch Fast Enough
Anthropic's Project Glasswing has found over 10,000 critical vulnerabilities in a month, but patching them fast enough is now the hard problem
Takeaways
- 10,000+ critical bugs found in one month: Anthropic's Project Glasswing and ~50 partners used Claude Mythos Preview to find over 10,000 high/critical vulnerabilities in critical software.
- Patching is now the bottleneck: Of 530 disclosed critical bugs, only 75 have been patched; some open-source maintainers have asked Anthropic to slow down disclosures.
- Benchmark-topping exploit capabilities: On ExploitGym (898 real CVEs), Mythos Preview successfully built working exploits for 157 instances vs. GPT-5.5's 120; only these two models showed meaningful kernel exploitation ability.
- OpenAI fires back: GPT-5.5-Cyber launched as a direct competitive response, with a broader access model via Trusted Access for Cyber; independent tests put the two models at rough parity.
- Glasswing expands to 150 more orgs: New partners span 15+ countries and now include power, water, healthcare, NATO, ENISA, Samsung, and SK Telecom.
- Claude Security now in public beta: Claude Security for Enterprise customers has already patched 2,100+ vulnerabilities using Claude Opus 4.7 in its first three weeks.
Project Glasswing, Anthropic's collaborative AI cybersecurity initiative, has just published its first major progress report -- and the numbers are striking. Since launching, Anthropic and its approximately 50 partners have used Claude Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities across the most systemically important software in the world. The finding that stands out most, though, is not the volume of bugs found. It's what's now bottlenecking the whole effort: humans can't patch fast enough.
Progress on software security used to be limited by how quickly we could find new vulnerabilities. Now it's limited by how quickly we can verify, disclose, and patch the large numbers of vulnerabilities found by AI. That's a fundamental shift in the economics of cybersecurity -- and it has implications for every organization that ships software.
What Claude Mythos Preview Actually Is
Claude Mythos Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. This is not a fine-tuned security scanner. It is a general reasoning model that happens to be extraordinarily good at understanding code at a deep, adversarial level.
Internal testing revealed the model's capacity to locate vulnerabilities and, crucially, combine those primitives into complete, end-to-end attack chains, a level of sophistication previously unseen in large language models. An "exploit primitive" is a building block of an attack -- for example, a memory corruption trick that lets you write to an arbitrary address. Chaining these into a full attack is what separates a theoretical bug from a working weapon.
The Numbers Behind the First Month
After one month, most partners have each found hundreds of critical- or high-severity vulnerabilities in their software. Several have told us that their rate of bug-finding has increased by more than a factor of ten. Some specific highlights from partners and external evaluators:
- Cloudflare found 2,000 bugs (400 of which are high- or critical-severity) across their critical-path systems, with a false positive rate that Cloudflare's team considers better than human testers.
- Mozilla found and fixed 271 vulnerabilities in Firefox 150 while testing Mythos Preview -- over ten times more than they found in Firefox 148 with Claude Opus 4.6.
- The UK's AI Security Institute (AISI) found that Mythos Preview is the first model to fully solve its rigorous corporate network attack simulations, completing an average of 22 out of 32 steps and achieving full end-to-end compromise in 30% of attempts.
- The latest Palo Alto Networks release included over five times as many patches as usual.

For the last few months, Anthropic has used Mythos Preview to scan more than 1,000 open-source projects, which collectively underpin much of the internet. So far, Mythos Preview has found what it estimates are 6,202 high- or critical-severity vulnerabilities in these projects (out of 23,019 in total). Of the subset that has been independently reviewed, 90.6% proved to be valid true positives, and 62.4% were confirmed as either high- or critical-severity.
The New Benchmarks Rewriting the Threat Model
To measure Mythos's capabilities rigorously, Anthropic supported the creation of two new academic benchmarks: ExploitBench (from Carnegie Mellon and Bugcrowd) and ExploitGym (from Berkeley RDI). These go far beyond traditional vulnerability scanners -- they test whether a model can build a working exploit from scratch.
ExploitGym is a benchmark of 898 real-world vulnerabilities spanning userspace programs, Google's V8 JavaScript engine, and the Linux kernel. Given a vulnerability and a proof-of-concept input, AI agents are tasked with crafting a full exploit that achieves unauthorized code execution. Anthropic's Claude Mythos Preview successfully exploited 157 of those 898 instances, and OpenAI's GPT-5.5 exploited 120.
Kernel exploitation was the sharpest dividing line: only Claude Mythos Preview (12 successes) and GPT-5.5 (22 successes) showed meaningful capability, while no other model managed more than one. On ExploitBench, which targets Chrome's V8 engine specifically, Mythos Preview reached Tier 1 (full arbitrary code execution) on 21 of 41 CVEs (51%). GPT-5.5 is the only other model to crack Tier 1, on just 2 CVEs.
One case study stands out. In one instance, Mythos Preview was able to create a near-deterministic exploit for a bug (CVE-2023-6702) where publicly known exploits were probabilistic and uncontrolled. Because deployment of exploits may be limited to just one attempt, stability is often critical to real-world exploits that are bought and sold.
The Patching Bottleneck Is the Real Story
Finding bugs is now the easy part. The hard part -- the part that hasn't scaled -- is everything that comes after. Anthropic's open-source scanning pipeline illustrates this vividly. Of the 530 high- or critical-severity bugs disclosed to maintainers so far, only 75 have been patched, and 65 of those have been given public advisories. The rest are stuck in a queue.
Several maintainers have told Anthropic they're currently severely capacity constrained, and some have even asked Anthropic to slow down its rate of disclosures because they need more time to design patches. On average, a high- or critical-severity bug found by Mythos Preview takes two weeks to patch. That two-week window, multiplied across thousands of vulnerabilities, is a significant exposure window.
The relative ease of finding vulnerabilities compared with the difficulty of fixing them amounts to a major challenge for cybersecurity. Confronting this challenge successfully will make our software far safer than before. Anthropic's advice to the industry right now:
- Software developers should shorten patch cycles and make security fixes available as quickly as possible, and make it easier for users to install updates.
- Network defenders should shorten patch testing and deployment timelines, enforce multi-factor authentication, and keep comprehensive logs for detection and response.
What Anthropic Is Shipping for Everyone Else
Not everyone gets access to Mythos Preview -- and Anthropic is explicit that it won't release it publicly until stronger safeguards exist. But the initiative has already produced tools for the broader market:
- Claude Security, released in public beta for Claude Enterprise customers, helps teams scan their codebases for vulnerabilities and generate proposed fixes. In the three weeks since launch, Claude Opus 4.7 has been used to patch over 2,100 vulnerabilities.
- The Cyber Verification Program now permits vetted security professionals to use Anthropic's models without standard cyber safeguards for legitimate penetration testing and red-teaming operations.
- Anthropic has also released the skills (custom instructions for repeated work), a codebase-scanning harness with subagents, and a threat model builder that maps a codebase to identify potential attack targets.
An Arms Race Is Already Underway
Anthropic's cautious, coalition-based rollout has triggered a direct competitive response. OpenAI CEO Sam Altman announced GPT-5.5-Cyber, a frontier AI model built specifically for cybersecurity, with a rollout beginning immediately to what the company is calling "critical cyber defenders." The move is a direct response to Anthropic's Claude Mythos Preview.
The two labs have settled into different access philosophies, with OpenAI casting a wider net through its Trusted Access for Cyber program and Anthropic running a tighter consortium. Independent testing puts GPT-5.5 and Mythos at rough parity, with sources telling reporters the two are "roughly on par" and one recent test putting Mythos narrowly ahead. The practical difference for most security teams is simply which program their organization can get approved for.
Anthropic has since significantly broadened Project Glasswing, extending access to Claude Mythos Preview to approximately 150 new organizations. The newly onboarded organizations span more than fifteen countries, with industries that were underrepresented in the initial cohort -- including power, water, healthcare, communications, and hardware -- now integrated into the program. The FT also reported that organizations given access include Okta, Samsung, SK Hynix, SK Telecom, NATO, and the EU's cybersecurity agency ENISA.
Who Wins, Who Loses, and What Comes Next
The clearest winners right now are organizations already inside the Glasswing perimeter -- they are getting a 10x improvement in vulnerability discovery before attackers do. Open-source maintainers are in a more ambiguous position: they're getting bugs reported faster than they can handle them, which is both a gift and a burden.
The speed of AI progress means that models as capable as Mythos Preview will soon be developed by many different AI companies. At present, no company -- including Anthropic -- has developed safeguards strong enough to prevent such models from being misused and potentially causing severe harm. That is why Anthropic has yet to release Mythos-class models to the public.
Anthropic's assessment is stark: within 6 to 12 months, many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse. In that world, cyberattacks could occur much more often, and in much more unpredictable forms. The window to harden critical infrastructure before that happens is exactly what Project Glasswing is racing to use.
The deeper question is whether the patching ecosystem -- open-source maintainers, enterprise security teams, patch deployment pipelines -- can be restructured fast enough to absorb what AI can now find. Anthropic's partnership with the Open Source Security Foundation's Alpha-Omega project is a start. But the scale of the problem -- thousands of confirmed critical bugs, a two-week average patch time, and overloaded maintainers -- suggests the industry has a structural gap to close, not just a backlog to clear.