Anthropic’s ‘Mythos’ AI Finds Deep Security Flaws

Anthropic Unveils ‘Mythos,’ an AI Too Powerful for Public Release

Artificial intelligence company Anthropic has revealed details about its most advanced AI model yet, named Claude Mythos. This new AI is so capable, especially in finding software security weaknesses, that Anthropic has decided it is too dangerous to release to the public. This marks a significant moment in AI development, shifting the focus from simply building powerful tools to urgently addressing their safety.

Mythos Surpasses Existing AI Models

Anthropic organizes its AI models into different tiers. Haiku is the cheapest and fastest, Sonnet is the standard workhorse model, and Opus is the most intelligent model currently available. Mythos, however, sits above all of these. It represents a completely new level of AI capability.

To show how advanced Mythos is, Anthropic shared benchmark results. On the SWE-bench verified test, which measures real-world software engineering skills, Mythos scored an incredible 93.9%. For comparison, Claude Opus, the best model people can use now, scored 80%. While this might seem like a small jump, AI experts know that even a few percentage points can mean a huge leap in performance.

Mythos also excelled on the Terminal-bench, a test for complex coding tasks in command-line environments. It achieved 82%, significantly higher than Opus’s 65.4%. These results suggest that AI models can continue to improve dramatically as they get larger, defying earlier predictions that progress might slow down without new AI designs.

Unprecedented Security Vulnerability Discovery

The most striking capability of Mythos is its ability to find security flaws. It recently discovered a 27-year-old vulnerability in OpenBSD, a highly secure operating system used for critical infrastructure like firewalls. This flaw allowed an attacker to remotely crash any machine running the system. Mythos also found a 16-year-old vulnerability in FFmpeg, a software library used in countless applications.

This ability to find long-hidden vulnerabilities is what makes Mythos so concerning. Anthropic itself has seen AI models being used for malicious purposes. In September 2025, they detected a Chinese state-sponsored hacking group using their Claude AI to infiltrate about 30 organizations, including tech companies and financial institutions. The AI handled 80-90% of the operation on its own. This was with Opus, a model that is less capable than Mythos.

Project Glasswing: A New Approach to AI Safety

Because of the risks, Anthropic has launched Project Glasswing. This initiative brings together leading companies to improve cybersecurity defenses before more dangerous AI capabilities become widely available. Anthropic plans to give these partner companies early access to Mythos. They will use the AI to find vulnerabilities in their own systems, allowing them to fix them before attackers can exploit them.

Anthropic is offering up to $100 million in usage credits to these companies. This project highlights a major shift: instead of building and releasing AI tools, companies must now focus on making them safe. Anthropic has stated that Mythos will not be released to the public. This is a policy decision, not a delay.

Why This Matters

The development of Mythos signals a critical turning point for artificial intelligence. For years, the AI industry followed a model: build a tool, release it, and let the market figure out its use. Mythos breaks this pattern because it is too powerful and potentially too dangerous for general release. Its creators have decided the public should not have access because it works *too* well, particularly in discovering security weaknesses.

This raises urgent questions about the future of AI and cybersecurity. If advanced AIs can find vulnerabilities faster than humans can fix them, how will we protect our digital world? The potential for misuse by malicious actors, including hackers and even nations, is a serious concern. Project Glasswing is Anthropic’s attempt to get ahead of these threats by using AI for defense.

The Future of AI Development and Cost

Anthropic is working to make advanced models more efficient. However, they admit the cost to run a model like Mythos is currently very high, making public access impractical for now. Prediction markets estimate a public launch, if it ever happens, could be years away, likely tied to safety evaluations rather than a release schedule.

The high cost of developing and running powerful AIs like Mythos might also mean that the most advanced tools will remain behind expensive paywalls. This could create a divide between those who can afford cutting-edge AI and those who cannot.

A New Era for AI

Some have compared Mythos to OpenAI’s GPT-2 release in 2019, which was also initially deemed too dangerous for public access. However, Anthropic argues this is different. GPT-2 was a language model capable of generating text, while Mythos can actively find and exploit security flaws. The consequences of Mythos’s capabilities are far more immediate and potentially damaging.

Anthropic believes their approach with Mythos and Project Glasswing will set a precedent for future AI releases. Companies may need to test models internally, brief governments, provide them to security experts for defense, and only then consider a general release once safeguards are in place. This marks a new era where AI companies are asking, “Is this safe enough?” rather than just “Is it good enough?” The gap between the most advanced AI systems inside research labs and what is available to the public is widening, and Mythos is a prime example of this growing divide.

Source: Claude Mythos Explained: Anthropic’s Most Dangerous Model Yet (YouTube)