Claude Mythos: Il modello IA troppo pericoloso per essere rilasciato

The Announcement

On April 8, 2026, Anthropic unveiled Claude Mythos Preview, described as their most powerful AI model ever built. Unlike every other Claude release, Mythos is not available to the general public. Instead, access is carefully restricted to security researchers and industry partners through Project Glasswing, a cross-industry cybersecurity coalition including AWS, Apple, Microsoft, Google, NVIDIA, Cisco, CrowdStrike, JPMorganChase, and the Linux Foundation.

This isn't a product launch. It's an insurance policy.

What Claude Mythos Can Do

The capabilities revealed in Mythos's system card and benchmarks are extraordinary—and that's precisely why Anthropic isn't releasing it to consumers.

Zero-Day Autonomy

In controlled red-team exercises, Claude Mythos autonomously discovered zero-day vulnerabilities in every major operating system:

• OpenBSD: A 27-year-old vulnerability previously unknown to the security community
• FFmpeg: A 16-year-old critical vulnerability
• Linux kernel: Privilege escalation vulnerability

This wasn't scripted. Mythos independently analyzed code, identified exploitable weaknesses, and documented them in formats exploit developers could immediately weaponize.

Benchmark Dominance

The performance gap between Mythos and Opus 4.6 is not incremental—it's categorical:

Opus 4.6 already represents the frontier of production AI. Mythos leaves it in the dust across every measured domain.

Why Project Glasswing Exists

Anthropic's decision to create Glasswing—and to not release Mythos publicly—reflects a maturation in how AI labs think about safety, deployment, and risk.

The Logic

Anthropic could monetize Mythos immediately. Early access pricing ($25/$125 per million input/output tokens after the preview period) would be highly profitable. Instead, they chose to:

1. Restrict access to vetted security researchers and industry partners
2. Front-load red-teaming through the Glasswing consortium rather than post-deployment incident response
3. Sequester the model while safeguards are developed
4. Promise that future Opus releases will receive these safeguards first, establishing a precedent that capability leads safety, not vice versa

This is the opposite of the industry norm: move fast, scale broadly, patch vulnerabilities as they're discovered in production. Anthropic is saying: Mythos is too dangerous to move fast with. Test it thoroughly. Then gate it.

The Financial Commitment

• $100M in usage credits committed to Glasswing partners
• $4M donated to open-source security organizations

These aren't marketing expenses. This is the cost of responsible disclosure at scale.

What This Means for Developers

For Most: Nothing Changes (Yet)

If you use Claude API, Bedrock, or Vertex AI today, you're on Opus 4.6 or earlier models. Claude Mythos won't be in your production stack—not yet, possibly never in its current form.

The safeguards being developed now will eventually ship in future Opus releases. That means better performance, but also more defensive AI systems—harder to jailbreak, less likely to autonomously discover and weaponize vulnerabilities.

For Security Researchers: Access Unlocked

If you're part of Glasswing or approved by the consortium, Mythos becomes available on:

• Claude API (Anthropic)
• Amazon Bedrock
• Google Vertex AI
• Microsoft Foundry

This effectively turns 50+ organizations into a coordinated red-team, all stress-testing the same model simultaneously, reporting vulnerabilities through structured channels.

The Precedent

This model of "capability gating + consortium red-teaming + safeguard development + controlled release" is new. If it works, expect other labs to adopt similar patterns. If it fails—if Mythos leaks, if vulnerabilities slip through Glasswing review—the entire industry learns that capability restrictions don't hold, and reverts to open deployment.

The Broader Context

Model Commoditization Continues

Meanwhile, open-source alternatives are accelerating:

• Nemotron 3 Super 120B (NVIDIA): Free, 7x faster via selective rounding and Mamba layers
• GLM 5.1 (Z.ai): 754B parameters, MIT-licensed, already on HuggingFace

The paradox: As frontier capabilities become restricted, open-weight alternatives gain traction. Anthropic's decision to gate Mythos may actually accelerate demand for open models, even if they're less capable.

Sam Altman's 2028 Timeline

OpenAI's Sam Altman recently stated he expects "automated researcher" capability by March 2028—AI that can independently conduct novel scientific research, run experiments, and publish findings. If Mythos is a proof-of-concept for that timeline, Altman's prediction just became more credible.

Key Takeaways

1. Mythos exists and works. Anthropic didn't announce vaporware. The benchmarks and vulnerability discoveries are real.

2. The lab made a risk calculation. Releasing it would be profitable in Q2 2026. Not releasing it reduces existential risk in the 2026-2028 window when autonomous AI capabilities mature rapidly.

3. This is a new deployment pattern. Restricted access + consortium red-teaming + staged safeguards + eventual public release (if safe) could become the standard for frontier models.

4. First-mover advantage matters, but so does not being the lab that caused the first major AI-discovered vulnerability to be weaponized. Anthropic is betting that reputation for safety is worth more than Q2 revenue.

5. The real story isn't about Mythos. It's about what happens next. What vulnerabilities does Glasswing find? How long until safeguards are deemed sufficient? What does the next Opus release look like?