Claude Mythos Preview and Project Glasswing Explained

Claude Mythos Preview is Anthropic's most powerful AI model ever built, announced on April 7, 2026 as part of Project Glasswing. The model has autonomously discovered thousands of zero-day vulnerabilities across every major operating system and web browser — including bugs that went undetected for up to 27 years — and Anthropic considers it too dangerous to release publicly.

What Is Claude Mythos Preview?

Claude Mythos Preview is a general-purpose language model that achieves state-of-the-art results across coding, reasoning, and computer use benchmarks — but its most striking capability is in cybersecurity. According to Anthropic's announcement, the model identified nearly all of its vulnerability discoveries "entirely autonomously, without any human steering."

Unlike previous Claude models, Mythos Preview can not only find vulnerabilities but also develop working exploits. In testing against Firefox 147's JavaScript engine, Opus 4.6 produced only 2 working exploits across several hundred attempts. Mythos Preview developed 181 working exploits and achieved register control on 29 more.

As of April 2026, Anthropic has stated it does not plan to make Mythos Preview generally available in its current form. The company plans to incorporate improved safety mechanisms into future Claude Opus releases before deploying Mythos-class capabilities at scale.

Benchmark Performance vs Opus 4.6

Mythos Preview outperforms Claude Opus 4.6 across every published benchmark, with the largest gains in cybersecurity and agentic coding tasks.

The most dramatic improvement is on SWE-bench Multimodal, where Mythos more than doubles Opus 4.6's score. The model also shows a 4.9x token efficiency improvement on BrowseComp, indicating substantially better computer use and search capabilities.

Zero-Day Vulnerabilities Discovered

Mythos Preview has autonomously discovered thousands of high- and critical-severity zero-day vulnerabilities across major software. As of April 2026, over 99% of the vulnerabilities found have not yet been patched, according to Anthropic's disclosure.

Notable Discoveries

OpenBSD — 27-year-old SACK vulnerability: A TCP sequence number handling flaw exploiting a signed integer overflow, enabling remote denial-of-service against any responding OpenBSD host. This bug existed undetected for 27 years.

FFmpeg H.264 codec — 16-year-old: A slice numbering collision with a sentinel value allowing an out-of-bounds heap write from specially crafted video frames. Automated fuzzing tools had hit this code path over 5 million times without triggering the bug.

FreeBSD NFS — 17-year-old (CVE-2026-4747): A remote unauthenticated root access vulnerability via stack buffer overflow in RPCSEC_GSS authentication. Mythos developed a working ROP chain exploit split across multiple packets — entirely autonomously.

Exploit Development at Scale

When tested against the OSS-Fuzz repository (~7,000 entry points across ~1,000 repositories), Sonnet 4.6 and Opus 4.6 each found 150-175 tier-1 crashes with only 1 tier-3 result each. Mythos Preview found 595 tier-1/tier-2 crashes, results at tiers 3-4, and 10 tier-5 findings — full control flow hijacks.

Anthropic reports that Mythos developed exploits "in hours that expert penetration testers said would have taken them weeks." The model also demonstrated the ability to reverse-engineer source code from stripped binaries, then identify vulnerabilities in closed-source software.

What Is Project Glasswing?

Project Glasswing is Anthropic's cybersecurity initiative that provides controlled access to Claude Mythos Preview for defensive security work. The program launched on April 7, 2026 with 12 partner organizations and approximately 40 additional critical infrastructure maintainers.

Launch Partners

The 12 founding Glasswing partners are: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and an additional undisclosed partner. Roughly 40 more organizations responsible for building or maintaining critical software infrastructure also received access.

Financial Commitments

Anthropic has committed significant resources to the initiative:

• $100 million in usage credits for Project Glasswing participants
• $2.5 million to Alpha-Omega and OpenSSF through the Linux Foundation
• $1.5 million to the Apache Software Foundation

Responsible Disclosure

Anthropic follows a 90+45 day coordinated vulnerability disclosure process and uses SHA-3 cryptographic commitments to prove vulnerability possession before patches are deployed. The company has contracted professional security contractors to manually validate every bug report before vendor notification.

Pricing and Availability

Claude Mythos Preview is not available to the general public. As of April 2026, access is restricted to Project Glasswing participants and approved critical infrastructure organizations.

For comparison, Claude Opus 4.6 is priced at $15/$75 per million input/output tokens. Mythos Preview costs approximately 67% more, reflecting its significantly higher capabilities.

What This Means for OpenClaw Operators

Mythos Preview is not directly accessible to OpenClaw operators today, but its existence signals several important developments for the AI agent ecosystem.

Security scanning will become standard. As Mythos-class models become more widely available, expect automated vulnerability scanning to become a routine part of software deployment — including for self-hosted AI agents like OpenClaw.

API costs for frontier models continue rising. At $25/$125 per million tokens, Mythos is the most expensive Claude model yet. Operators running cost-sensitive workloads should continue using cost-optimized model routing strategies.

Security hardening matters more than ever. If AI models can autonomously find and exploit decades-old vulnerabilities, every self-hosted deployment needs proper security hardening. Review your firewall rules, API key storage, and network exposure.

Limitations and Concerns

Despite its capabilities, Claude Mythos Preview raises significant concerns that Anthropic has acknowledged.

Dual-use risk. The same capabilities that find defensive vulnerabilities could be used offensively. This is precisely why Anthropic has restricted access and does not plan a public release of Mythos Preview in its current form.

Not a replacement for human security teams. While Mythos can find and exploit vulnerabilities autonomously, interpreting business impact, prioritizing remediation, and managing disclosure still requires human judgment.

Cost prohibitive for most. At $25/$125 per million tokens, comprehensive scanning runs cost $10,000-$20,000 depending on codebase size. Individual vulnerability hunts run under $50-$2,000, but this is still far beyond typical operator budgets.

Unpatched vulnerabilities. Over 99% of discovered vulnerabilities remain unpatched as of the announcement. The coordinated disclosure process means months will pass before fixes reach end users.

Future safeguards uncertain. Anthropic has stated it will launch new safeguards with an upcoming Claude Opus model, but the specifics and timeline remain unclear.

Related Guides

• AI Agent Security Risks: What You Need to Know
• OpenClaw 3-Tier Security Hardening
• Claude Opus 4.6 OpenClaw Guide
• Complete Guide to OpenClaw

Frequently Asked Questions

What is Claude Mythos Preview?

Claude Mythos Preview is Anthropic's most powerful unreleased AI model, announced in April 2026. It is a general-purpose model with unprecedented cybersecurity capabilities, able to autonomously find and exploit zero-day vulnerabilities in major operating systems and browsers.

Is Claude Mythos available to the public?

No. As of April 2026, Claude Mythos Preview is not publicly available. Anthropic has restricted access to Project Glasswing partner organizations and approximately 40 additional organizations that maintain critical infrastructure. Anthropic has stated it does not plan to make Mythos Preview generally available.

What is Project Glasswing?

Project Glasswing is Anthropic's cybersecurity initiative that gives selected organizations access to Claude Mythos Preview for defensive security work. Launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and Palo Alto Networks.

How much does Claude Mythos cost?

For Project Glasswing participants, Claude Mythos Preview is priced at $25 per million input tokens and $125 per million output tokens. Anthropic has also committed $100 million in usage credits for Glasswing participants.

How does Claude Mythos compare to Claude Opus 4.6?

Mythos Preview significantly outperforms Opus 4.6 across all benchmarks. On CyberGym it scores 83.1% vs 66.6%, on SWE-bench Verified 93.9% vs 80.8%, and on exploit development Mythos achieved 181 working Firefox exploits where Opus 4.6 managed only 2.