Anthropic Keeps Mythos Under Wraps — The Most Dangerous AI Model Yet?

Anthropic has done something unusual. In early April 2026, the company announced its latest AI model, Mythos — then promptly refused to release it to the public. The last time a major AI developer withheld a model over safety concerns was OpenAI's GPT-2 in 2019. That was seven years ago. The stakes this time are considerably higher.

Mythos isn't just another large language model. According to Anthropic's own 245-page technical report, the model operates at the level of a senior software engineer, capable of spotting subtle bugs, self-correcting mistakes, and exploiting vulnerabilities across every major operating system and web browser. In testing, it found critical security flaws in all of them. 99 percent of those vulnerabilities have not yet been patched.

What Makes Mythos Different?

The model scored 31 percentage points higher than Anthropic's previous flagship, Opus 4.6, on the USAMO 2026 Mathematical Olympiad — a grueling, two-day proof-based competition. But it's the cybersecurity capabilities that have regulators on edge.

The UK's AI Security Institute (AISI), which received early access to Mythos, found that the model succeeded in expert-level hacking tasks 73 percent of the time. To put that in perspective: prior to April 2025, no AI model could complete those tasks at all. In less than a year, we've gone from zero to a 73 percent success rate on tasks that require deep system-level understanding.

Anthropic claims Mythos can outstrip all but the most skilled human hackers at identifying and exploiting software vulnerabilities. The model found critical faults in every widely used operating system and web browser during testing, and the company has only disclosed a fraction of what it says it has found.

Project Glasswing — Limited Access for the Big Players

Instead of a public rollout, Anthropic is limiting access through an initiative called Project Glasswing. The program gives a select group of organizations access to Mythos for defensive cybersecurity work — scanning their own networks and patching problems before vulnerabilities become public knowledge.

The initial cohort reads like a who's-who of tech: Microsoft, Google, Apple, Amazon Web Services, JPMorgan Chase, and Nvidia. These companies get to use one of the most powerful hacking tools ever created to fix their own infrastructure. Everyone else? They're waiting for patches

It's a logical approach. Let the companies with the most at stake and the most sophisticated security teams use the model to harden the systems that billions of people rely on. But it also raises uncomfortable questions about who gets access to defensive AI capabilities and who gets left behind.

Not Everyone Is Buying the Alarm

Cybersecurity experts are divided on whether Mythos represents a genuine breakthrough or an expected step along a trajectory we've been on for years. The model's capabilities are real — there's no debate about that. But some researchers argue that Anthropic's framing may be as strategic as it is scientific.

Positioning Mythos as "too dangerous to release" reinforces Anthropic's brand as the safety-first AI company, especially as its rival OpenAI faces a high-profile trial over its founding mission. And with Anthropic recently expanding into defense contracts through its Pentagon partnerships, the line between responsible disclosure and strategic positioning gets blurry.

What This Means for Developers and Security Teams

For developers and security professionals, Mythos is a wake-up call. The pace of AI capability advancement in cybersecurity is accelerating faster than most organizations can adapt. If a model can find 99 percent critical vulnerabilities across every major platform, the window for proactive defense is narrowing.

The practical takeaway: security-first development isn't optional anymore. AI-assisted code generation is already widespread (tools like Claude and Cursor are everyday tools for millions of developers), and the same models that help you write code can now find the flaws in it at a level that rivals expert human penetration testers.

The Road Ahead

Anthropic has said it will continue evaluating Mythos through Project Glasswing before making any decisions about broader access. For now, the model remains a preview of what's coming — a glimpse of an AI capability landscape where the gap between offensive and defensive applications shrinks every quarter.

Whether Mythos is a genuinely dangerous breakthrough or a carefully managed narrative, one thing is clear: AI-powered cybersecurity is entering a new phase. The models are here. The vulnerabilities are being found. The only question is who gets to use the tools first.

Sources:

• Scientific American — "What is Mythos, Anthropic's unreleased AI model, and how worried should we be?" https://www.scientificamerican.com/article/what-is-mythos-and-why-are-experts-worried-about-anthropics-ai-model/
• Anthropic — Mythos Technical Report https://cdn.sanity.io/files/4zrzovbb/website/7624816413e9b4d2e3ba620c5a5e091b98b190a5.pdf
• UK AI Security Institute — Evaluation of Claude Mythos https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
• Anthropic — Project Glasswing https://www.anthropic.com/glasswing
• The Verge — "DeepMind paper argues LLMs will never be conscious" https://www.theverge.com/ai-artificial-intelligence/917755/musk-altman-openai-xai-gossip