Claude Mythos Preview Launch: Anthropic's Most Powerful Model Targets Zero-Day Vulnerabilities via Project Glasswing

The Most Capable Model Anthropic Has Ever Built — Deployed Exclusively for Defense

On April 7, 2026, Anthropic made one of the most consequential AI announcements in recent memory. The company officially unveiled Claude Mythos Preview, a model it describes as "by far the most powerful AI" it has ever developed — and immediately restricted it from general public access.

Instead of a standard API rollout, Anthropic launched Project Glasswing, a tightly controlled cybersecurity initiative giving select organizations access to Mythos Preview exclusively for defensive security work. The decision reflects both the model's extraordinary capabilities and the serious dual-use risks those capabilities introduce.

What Makes Mythos Preview Different

Mythos Preview is not just an incremental upgrade over Claude Opus 4.6. Anthropic's own internal testing reveals a capability discontinuity unlike any previous model generation.

Autonomous Exploit Development at Scale

The most striking benchmark comes from direct comparison with Opus 4.6 on the same vulnerability corpus. When tested against Firefox-level challenges, Opus 4.6 produced 2 successful working exploits from several hundred attempts. Mythos Preview produced 181 working exploits — an approximately 90x improvement.

This leap extends across operating systems and browsers. Tested against the OSS-Fuzz corpus of roughly 7,000 entry points, both Sonnet 4.6 and Opus 4.6 achieved a complete control-flow hijack (tier 5) only once each. Mythos Preview achieved tier 5 on 10 separate targets, plus hundreds of lower-severity crashes.

Zero-Day Discovery at Unprecedented Depth

Over the past few weeks before launch, Anthropic used Mythos Preview to scan major operating systems and web browsers, identifying thousands of previously unknown critical and high-severity vulnerabilities. The company reports that over 99% of these findings have not yet been patched, limiting public disclosure.

A representative sample of confirmed discoveries:

Of 198 manually reviewed vulnerability reports, 98% of Mythos's severity assessments were within one severity level of human expert assessments — a level of accuracy that rivals experienced security researchers.

Cost Efficiency at Scale

Beyond raw capability, the economics of Mythos-driven security research are striking:

• 1,000 runs through the OpenBSD exploit scaffold: under $20,000
• A single successful OpenBSD exploit: under $50
• Full FFmpeg vulnerability discovery pass: roughly $10,000
• Individual Linux kernel exploits: $1,000–$2,000

At these price points, defensive teams can now run continuous, large-scale automated vulnerability discovery against their own codebases at a fraction of traditional red team costs.

Project Glasswing: Controlled Deployment for Defense

Recognizing the dual-use risk — the same capabilities that find vulnerabilities can also exploit them — Anthropic chose not to offer Mythos Preview as a standard commercial product. Project Glasswing is the alternative.

The initiative currently includes approximately 50 organizations across technology, finance, and critical infrastructure sectors, among them Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and Nvidia. Each participant uses Mythos Preview for defensive security work and shares learnings with the broader industry.

Access is available via the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry, priced at $25 per million input tokens and $125 per million output tokens — substantially higher than standard Claude pricing, reflecting both the model's power and the restricted access model.

Anthropics' overarching goal is to coordinate a global effort to patch critical infrastructure vulnerabilities before adversaries — state actors, criminal groups, or others — can replicate Mythos-level capabilities on their own.

The Dual-Use Dilemma

Anthropichas been transparent about the risk calculus. The cybersecurity capabilities in Mythos Preview were not deliberately trained — they emerged as a downstream consequence of general improvements in code understanding, reasoning, and agentic autonomy. That framing carries a sobering implication: future, more capable models will almost certainly be even more dangerous in the wrong hands.

This announcement has drawn significant regulatory attention. Reports indicate that the US Treasury Secretary and Federal Reserve Chair have convened meetings with major US bank CEOs specifically over Mythos's implications for financial sector cybersecurity.

Anthropicacknowledges the tension explicitly: "the same capabilities that can bolster cyber defenses can also be weaponized by attackers." The company's bet is that controlled, coordinated deployment under Project Glasswing represents the most responsible path forward.

Usability and Access Considerations

For security professionals, the Project Glasswing access model means Mythos Preview is not a tool you can simply subscribe to and use tomorrow. Organizations must apply for participation, meet Anthropic's security and compliance criteria, and commit to sharing defensive findings.

The $25/$125 per-million-token pricing is roughly 8–10x the cost of standard Opus 4.6 access, positioning Mythos Preview firmly as an enterprise and critical infrastructure tool rather than a developer utility. For large organizations with substantial vulnerability management programs, however, the cost-per-exploit figures suggest it could rapidly pay for itself.

Pros and Cons

Strengths:

• Unprecedented autonomous exploit development capability (90x improvement over Opus 4.6)
• Discovers vulnerabilities in codebases that have resisted human and automated review for decades
• Favorable economics for large-scale defensive security programs
• Coordinated, responsible deployment model reduces immediate misuse risk
• Broad partner ecosystem accelerates patching of discovered vulnerabilities

Limitations:

• Not publicly accessible — requires Project Glasswing membership
• Premium pricing ($25/$125 per million tokens) limits accessibility to large organizations
• No guaranteed path to general availability
• Dual-use risk is real and acknowledged; broader proliferation of similar capabilities is inevitable
• Regulatory environment remains uncertain

Outlook: Redefining the Security Landscape

Claude Mythos Preview's launch marks a genuine inflection point in AI-assisted security research. The shift from "AI can sometimes help find bugs" to "AI can autonomously develop working exploits at industrial scale" changes the defensive calculus for every major software vendor and critical infrastructure operator.

Anthropicis urging organizations to take immediate practical steps: deploy frontier models like Opus 4.6 for their own vulnerability discovery now (while Mythos access is restricted), significantly reduce patch deployment timelines, and automate incident response pipelines.

The broader industry implication is harder to ignore. If Anthropic can build a model with these cybersecurity capabilities as an emergent byproduct of general capability improvements, other frontier labs will likely reach similar capability levels within months to a year. The window for coordinated defensive action through Project Glasswing may be narrower than it appears.

Conclusion

Claude Mythos Preview is the most consequential AI security announcement of 2026. It demonstrates that frontier AI models have crossed a threshold where autonomous, expert-level exploit development is no longer theoretical. Anthropic's decision to deploy it exclusively through the controlled Project Glasswing framework reflects both the gravity of that capability and a serious attempt to channel it defensively.

For security teams at large enterprises and critical infrastructure organizations, Project Glasswing membership is worth pursuing actively. For the broader industry, this announcement is a signal to dramatically accelerate patching and vulnerability management programs — because the AI-powered adversary landscape is arriving faster than most forecasts suggested.