AI News

Claude Mythos Preview: Anthropic builds an AI too powerful to be made public

Anthropic unveils Claude Mythos Preview, a model able to detect previously unknown security flaws. But it is reserved for a club of 40 tech giants. What this means for SMB cybersecurity and unequal access to the best technologies.

6 min read
AnthropicClaude MythosCybersécuritéIAPMEProject Glasswing
⚡ The news in 30 seconds

Anthropic builds an AI able to hack any software — and keeps it secret

On 7 April 2026, Anthropic unveiled Claude Mythos Preview, its most powerful AI model to date. Able to discover and exploit previously unknown vulnerabilities across all major operating systems, it found a 27-year-old flaw in OpenBSD and succeeded in 181 attacks on Firefox where its predecessor succeeded in only 2. The problem: this model is deemed too dangerous to be made public. Only 40 tech giants have access to it via Project Glasswing. An industry first that raises fundamental questions about unequal access to the most advanced technologies.

For SMBs, this is not just tech news. It is a warning sign about the concentration of AI super-powers in the hands of a small club — and about the cybersecurity risks that follow.

The hidden opportunity

While direct access to Mythos is impossible for SMBs, this announcement reveals important trends:

🛡️

Security awareness at the highest level

The fact that Anthropic locks down its own model shows just how critical cybersecurity has become. Companies that invest now in their AI security posture will be better prepared when Mythos's capabilities spread indirectly through the patches and updates of the major vendors.

🔍

The vulnerability of legacy software ecosystems

Mythos found 16- and 27-year-old flaws in software nonetheless reputed to be secure. This is a reminder that technical debt and legacy systems are sieves. SMBs that keep their software up to date and eliminate obsolete dependencies drastically reduce their attack surface.

⚖️

A precedent for AI regulation

This Anthropic decision could inspire the European AI Act and regulators worldwide. SMBs that anticipate compliance requirements and document their AI uses today will have a competitive advantage tomorrow over companies caught off guard.

The major risk

⚠️

An unprecedented digital divide

For the first time, an AI model of general interest — cybersecurity concerns everyone — is reserved for a club of large companies. This creates a fundamental imbalance: tech giants can discover and fix flaws before everyone else, while SMBs remain vulnerable.

🔓

The leak is inevitable

Historically, restrictive technologies eventually leak. When Mythos's capabilities spread — through theft, reverse engineering or mere rumor — malicious actors could have them before defenders do. SMBs must prepare for a world where attacks become more sophisticated without having access to the same defensive tools.

Our recommendation

In the face of this evolution, here are the priority actions for SMBs:

1

Strengthen your security monitoring immediately

Subscribe to the security alerts of your main software vendors. Put in place a patch management process with short deadlines: 48h for critical fixes, 1 week for important ones. Never leave a known flaw unpatched.

2

Audit your technical debt

Identify legacy systems, unmaintained software and obsolete dependencies in your infrastructure. Every component that is not updated is a potential target. Budget for a gradual modernization — the investment is lower than the cost of a breach.

3

Document and bring your AI uses into compliance

The European AI Act and the regulations that follow will require more and more transparency. Document which AI tools you use, for what purposes, and what data they process. This documentation will be your regulatory shield and a commercial asset.

In summary

The news
Anthropic launches Claude Mythos Preview, an ultra-high-performance cybersecurity AI reserved for 40 companies
The problem
Unequal access to the best defensive technologies, concentration of AI super-powers
Action 1
Strengthened security monitoring and accelerated patch management
Action 2
Audit of technical debt and modernization of legacy systems

Frequently asked questions

What exactly is Claude Mythos Preview?

It is the most advanced AI model ever developed by Anthropic. It excels at detecting software vulnerabilities, able to find and exploit zero-day flaws across all major operating systems and browsers. Unlike previous models, it is not available to the public.

Why does Anthropic not make this model public?

Anthropic considers the model too dangerous for broad release. Its autonomous hacking capabilities could be used by malicious actors to compromise critical systems. The company chose caution by limiting access to a consortium of large companies via Project Glasswing.

Which companies have access to Mythos?

About 40 major technology companies: AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and other critical infrastructure players. Anthropic offers 100 million dollars in usage credits to the consortium.

What is the impact for an SMB without access to Mythos?

SMBs remain dependent on the patches released by the major vendors. The concern is that Mythos could discover massive vulnerabilities in widely used software, creating a window of risk between discovery and the public patch. SMBs must strengthen their security monitoring and their responsiveness.

For technical profiles

Technical performance of Claude Mythos Preview:

MetricClaude Opus 4.6Claude Mythos PreviewImprovement factor
Successful Firefox exploits2 out of several hundred181 out of several hundred90x
Control flow hijack OSS-Fuzz010 targets patchedInfinite
Oldest flaw foundN/A27 years — OpenBSDN/A
AccessibilityPublicConsortium only-

Confirmed Project Glasswing members: AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks. Anthropic commits to 100 million dollars in usage credits for the consortium.

Implications for security research: This restriction approach raises the question of disclosure responsibility — when a tool can discover vulnerabilities at industrial scale, who should have access to it? The cybersecurity community is divided between those who welcome the caution and those who fear a concentration of discovery power.

Related articles