1300 166 138 Tell us what you need
News / Anthropic

Anthropic’s most powerful AI is too dangerous to release, and that tells us something

By Monika Tantau • Published 7 months ago
Anthropic Claude Mythos Preview — AI model too dangerous to release publicly

Key takeaways

• Anthropic released Claude Mythos Preview, which can find and exploit zero-day vulnerabilities in every major OS and browser
• Instead of a public release, Anthropic launched Project Glasswing, a consortium of 40 companies using Mythos to patch critical infrastructure
• The decision connects to documented performance degradation in Claude Code since February 2026, linked to changes in how the model handles internal reasoning
• Developers report an 80x increase in API requests as teams compensate for quality regression
• The story raises questions about the gap between frontier AI capability and production reliability at scale

Most Read

Anthropic has just unveiled its most capable AI model yet. Claude Mythos Preview can find and exploit hidden security vulnerabilities in every major operating system and web browser. In tests, it discovered a 27-year-old flaw in OpenBSD, a system known specifically for its security. It scores 93.9 percent on SWE-bench Verified, a standard benchmark for real-world coding tasks, and 97.6 percent on the 2026 US Mathematics Olympiad. Those are extraordinary numbers, but what happened next is what makes this story genuinely unusual.

Anthropic’s response to building something this capable was not to release it. Instead, the company launched Project Glasswing, a coordinated effort involving more than 40 technology companies including Apple, Amazon and Microsoft, to use Mythos Preview to find and patch critical software vulnerabilities before bad actors can exploit them. That decision sits at the centre of a broader story about what is actually happening inside AI labs right now.

It connects, in ways that are not entirely comfortable, to complaints from developers who say Claude Code, Anthropic’s AI coding assistant, has become noticeably worse in recent months. And it lands at a particularly charged moment. OpenAI recently closed a $122 billion funding round valuing the company at $852 billion, and is widely expected to announce major new capabilities of its own.

What Claude Mythos Preview Can Do and Why It Changes Cybersecurity

Anthropic published detailed test results alongside the Mythos announcement. The findings are unusually candid. Researchers found that Mythos Preview can identify and exploit security vulnerabilities across every major operating system and browser when directed to do so.

It can reverse-engineer software it has never seen before to find hidden flaws, and it can turn known-but-unpatched bugs into working attacks. These are capabilities typically associated with top-tier human security researchers.

Anthropic notes that over 99 percent of the vulnerabilities Mythos found during internal testing remain unpatched, which is why the company cannot disclose specifics. Even the small fraction it can discuss represents, in its own words, “a substantial leap” in what AI models can do in cybersecurity contexts.

The benchmark scores are impressive, but it is the cybersecurity capabilities that make this model genuinely different. Finding a 27-year-old previously unknown flaw in OpenBSD is not an abstract test result. It demonstrates that the model can work like a skilled security researcher, independently and at scale.

Anthropic withholding Claude Mythos from public release due to cybersecurity risks

Why Anthropic Is Withholding Claude Mythos from the Public

The official explanation is safety. Anthropic argues that a model capable of autonomously exploiting vulnerabilities across critical infrastructure represents a meaningful uplift in offensive cyber capability. Making it generally available before defenses are in place would be irresponsible.

That reasoning is credible. The model’s capabilities are genuinely novel, and the threat is concrete. A bad actor with access to Mythos-level offensive capability would have a significant asymmetric advantage over defenders.

But the safety explanation does not exist in a vacuum. Anthropic is also managing real infrastructure pressure. In early April, the company blocked Claude Pro and Max subscribers from using their plan limits with third-party tools like OpenClaw. The reason given: autonomous AI agent sessions were consuming six to eight times more computing resources per session than typical users. The infrastructure was not built for that level of demand.

Around the same time, developers began filing detailed complaints about Claude Code’s performance. The GitHub issue, filed by Stella Laurenzo and backed by hundreds of developer comments, documented a clear quality drop. It correlated precisely with a February 2026 update, specifically a change in how the model handles its internal reasoning process.

When that internal reasoning was restricted, output quality dropped noticeably. InfoWorld and several developer publications covered the story, highlighting the impact on enterprise teams who had built workflows around Claude Code.

Claude Code Performance Drop: The Thinking Content Problem

The connection matters. It suggests that the kind of deep reasoning that makes Mythos Preview so capable, the same quality on display in its benchmark results, comes at a significant computing cost. That cost is difficult to sustain across millions of everyday users.

Mythos Preview is almost certainly a more powerful reasoner than the model behind Claude Code. But more powerful reasoning requires more computing resources. If Anthropic is managing server load by limiting how much internal reasoning Claude Code does in real time, that directly affects the quality of the answers developers get.

This is not speculation. The GitHub issue documents an 80x increase in API requests from teams trying to compensate for quality regression by running more attempts on the same tasks. That creates a self-reinforcing infrastructure problem: quality degrades, developers make more requests, the load increases, quality degrades further.

Project Glasswing — Anthropic's consortium of 40 tech companies using Claude Mythos to patch vulnerabilities

Project Glasswing, Compute Limits, and the Competitive Race

Project Glasswing is, in part, a strategic response to this situation. By channelling Mythos Preview’s capabilities into a controlled group of 40 technology partners focused on defensive security, Anthropic can demonstrate what the model can do and establish a responsible-release narrative, all without exposing it to the open market before the infrastructure is ready to handle the demand.

This is not cynicism. The safety reasoning is real. A model that can autonomously exploit vulnerabilities across critical infrastructure probably should not be available to anyone with a subscription. But it would be naive to suggest that the compute constraint plays no role in the timeline.

The broader point is that Anthropic is navigating a situation that every frontier AI lab faces in some form: the gap between what the model can do in a controlled environment and what the infrastructure can reliably deliver at production scale. Mythos Preview represents a genuine leap in capability. Project Glasswing is, among other things, a way of managing that gap responsibly.

The competitive context matters here. OpenAI closed a $122 billion funding round in late March 2026, reaching a valuation of $852 billion, and is widely expected to make major model announcements of its own. Anthropic is managing infrastructure strain, a staged model rollout, and competitive pressure all at once.

Whether safety, computing limits, or competition is the main reason behind restricting Mythos Preview is probably the wrong question. In practice, all three are likely present. The more useful observation is that the AI industry has entered a phase where capability is advancing faster than the infrastructure needed to deliver it reliably and safely at scale.

What Claude Mythos Means for Australian Businesses Using AI Tools

For Australian organisations that have integrated Claude Code into their development workflows, the practical implications are immediate. The quality regression is documented and real. Running more iterations, adjusting prompts or switching to different models adds friction and cost.

The longer-term picture is more open-ended. If Anthropic’s infrastructure catches up with its model development, the tools available to Australian businesses will become significantly more powerful. Mythos Preview is a preview of that future. The question is timing, and whether OpenAI or another competitor gets there first.

For teams building AI workflows, this is a useful reminder that capability and reliability are not the same thing. Relying on a single provider for critical AI infrastructure carries real risk. Not because the technology is failing, but because it is moving faster than the systems designed to support it.

Australian businesses looking for AI automation solutions designed for reliability, not just raw capability, can explore how Aivy works as an AI automation agency, and learn more about how we approach AI chatbot development for enterprise contexts.

Scroll to Top