Claude Mythos: A Sub-$50 Run Found a 27-Year-Old OpenBSD Bug. Is Your Code Next?

Claude Mythos Preview is the Anthropic model that, when directed to, found and exploited zero-day vulnerabilities in every major operating system and web browser. Anthropic announced it on April 7, 2026 and kept it to Project Glasswing partners. The run that found a 27-year-old OpenBSD bug cost under $50. One month later partners reported more than 10,000 high or critical vulnerabilities.

TL;DR: Anthropic's Claude Mythos Preview finds and exploits zero-days at a pace no human team matches. Its April 2026 red team report covered every major operating system and browser. A month into Project Glasswing, partners had found more than 10,000 high or critical vulnerabilities. The model is not publicly available, but the gap it exposes is. Shrink patch windows, test internet-facing code first, track your open-source dependencies and audit code before it ships.

What Anthropic Reported in April 2026

Anthropic announced Claude Mythos Preview on April 7, 2026. Its Anthropic red team report says the model is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when a user directs it to. Anthropic chose not to make the model generally available and released it only to a limited group of partners through Project Glasswing.

The report's examples are specific:

  • OpenBSD. A 27-year-old bug in the TCP SACK handling that let an attacker crash any OpenBSD host that responds over TCP. It is now patched. OpenBSD is the operating system best known for security correctness, which is why this finding drew attention.
  • FFmpeg. A 16-year-old vulnerability in the H.264 codec, one of FFmpeg's most widely used. The underlying flaw traces to a 2003 commit. FFmpeg sits inside a very large share of the world's video software.
  • FreeBSD. A working NFS exploit that split a 20-gadget return-oriented programming chain across multiple packets.
  • Browsers. A browser exploit that wrote a complex JIT heap spray and escaped both the renderer sandbox and the operating system sandbox.

Anthropic also wrote that over 99% of the vulnerabilities it had found were not yet patched when it published the report. That figure describes April 2026. Maintainers have been fixing findings since then, so read it as a measure of the backlog at announcement, not as the state of the software today.

The Cost Numbers in Context

The most quoted figure is the cost. Anthropic ran its scaffold about a thousand times against OpenBSD and the total cost was under $20,000. The specific run that found the SACK bug cost under $50. That is the accurate reading of the "$50" headline: one successful run, identified in hindsight, out of a search that cost hundreds of times more.

Even read correctly, the economics are striking. A search that turns up a 27-year-old remote crash in a hardened operating system for under $20,000 is far cheaper than the expert time that kind of finding has historically needed. The defender's takeaway is not that every bug now costs $50. It is that the budget needed to find serious flaws in mature code has dropped sharply.

One Month of Project Glasswing

On May 22, 2026 Anthropic published a Project Glasswing update covering the first month of partner use. The headline numbers:

SourceReported finding
All Glasswing partnersMore than 10,000 high or critical vulnerabilities in one month, across roughly 50 partners
Cloudflare2,000 bugs in critical-path systems, 400 of them high or critical
Mozilla271 vulnerabilities found and fixed in Firefox 150, over ten times more than in Firefox 148 with Claude Opus 4.6
MicrosoftNo count given. Microsoft said its patch volume will keep trending larger for some time.
Open sourceMore than 1,000 projects scanned. Mythos Preview flagged 23,019 candidate vulnerabilities in total, 6,202 of them high or critical.

Anthropic checked a sample of the open-source results. Of 1,752 high or critical findings it assessed, 1,587 (90.6%) were valid and 1,094 (62.4%) were confirmed as high or critical. At that rate it projected nearly 3,900 confirmed high or critical open-source vulnerabilities. Several partners also reported that their bug-finding rate rose by more than a factor of ten.

Cloudflare is the number to sit with. It is a security company with a security-focused engineering team. A month of AI-assisted review still produced 400 serious findings. Code that has never been reviewed at all is unlikely to fare better.

Who Can Use Mythos

Mythos Preview was never released to the public. In May Anthropic said it had yet to release Mythos-class models to the public. On June 9, 2026 it introduced two variants of a newer model. Fable 5 is distributed widely with cybersecurity safeguards. Mythos 5 is the same model with fewer safeguards, deployed through Project Glasswing as an upgrade to Mythos Preview. Coverage of the launch, including Constellation Research, puts the price at $10 per million input tokens and $50 per million output tokens.

For defenders this access model cuts both ways. The strongest offensive capability sits with vetted partners. The general model still makes code review, triage and phishing faster for everyone, including attackers. Our analysis of Mythos 5 for incident response teams covers what that means for DFIR workflows.

What This Means for Your Organization

Nothing in these reports created new vulnerabilities in your code. They showed how much was already there and how cheaply it can now be found. Four practical consequences follow.

Patch windows have to shrink

The traditional cycle of test, stage, schedule and deploy over two to four weeks assumed that exploit development was slow. Faster discovery and faster exploit development shorten the time between disclosure and attack. Critical patches on internet-facing systems need a path measured in days, not weeks.

Annual testing leaves long gaps

A once-a-year assessment leaves eleven months in which new code and new disclosures go untested. Move high-risk applications to quarterly testing or to testing after every major release.

Unmaintained software is the weakest point

Software that no longer receives security updates will not get fixes for what AI-assisted review finds. Replace it or isolate it.

Open-source dependencies carry the backlog

The open-source numbers show where much of the risk lives. Keep a software bill of materials, following CISA's SBOM guidance, so you can answer quickly when a dependency you use turns up in a disclosure.

Why AI-Augmented Testing Matters Now

The same class of capability that makes Mythos dangerous makes modern testing effective. Our red team uses AI-augmented tooling alongside manual testing to probe applications, APIs, infrastructure and dependencies at the depth an adversary would. The difference is which side receives the findings. You get a report with prioritized findings and remediation guidance before anyone else does.

AI-assisted review is strong at volume and pattern. Business logic flaws, authorization mistakes and architecture weaknesses still need a human tester who understands what the application is supposed to do. The useful combination is both, guided by the NIST Cybersecurity Framework and checked against the CISA Known Exploited Vulnerabilities catalog.

What You Should Do This Week

  1. List your internet-facing applications and APIs. They are the first code anyone will analyze.
  2. Check your patch cadence for critical fixes. If it is measured in weeks, find out what would make it days.
  3. Build or refresh your SBOM. You cannot patch a dependency you do not know you use.
  4. Audit AI-generated code before it ships. Our AI code audit covers the patterns these models find.
  5. See what is visible from outside. Run the free external reconnaissance scan on our homepage, then order a penetration test for the systems that matter most.

References and Further Reading

Frequently Asked Questions

What is Claude Mythos and why is it a security concern?

Claude Mythos Preview is an Anthropic model that, when directed to, can find and exploit zero-day vulnerabilities in every major operating system and web browser. Its findings include a 27-year-old OpenBSD bug and a 16-year-old FFmpeg vulnerability. It lowers the cost of finding serious flaws in mature code.

Did a Mythos vulnerability really cost only $50 to find?

Only in a narrow sense. Anthropic ran its scaffold about a thousand times against OpenBSD for under $20,000 in total. The single run that found the 27-year-old SACK bug cost under $50.

How many vulnerabilities has Claude Mythos found?

Anthropic reported that Project Glasswing partners found more than 10,000 high or critical vulnerabilities in the first month, including 2,000 bugs at Cloudflare and 271 in Firefox 150. Across more than 1,000 open-source projects, Mythos Preview flagged 23,019 candidate vulnerabilities, 6,202 of them high or critical.

Are the Mythos vulnerabilities still unpatched?

Anthropic wrote in April 2026 that over 99% of the vulnerabilities it had found were not yet patched. That was the backlog at announcement. Many have been fixed since, including the OpenBSD bug.

Can I use Claude Mythos to scan my code?

Not directly. Mythos Preview was limited to Project Glasswing partners and Mythos 5 is deployed through the same program. Fable 5, released in June 2026, is broadly available with cybersecurity safeguards.

How can companies protect against AI-discovered vulnerabilities?

Shrink patch windows for critical fixes, test internet-facing code first and more often, replace unmaintained software, keep a software bill of materials and audit AI-generated code before it ships.

Make Sherlock Forensics your preferred source in Google Search