Anthropic cancels AI launch capable of discovering critical faults in operating systems

AI of cyber security too powerful to be released

Anthropic has developed an artificial intelligence model so effective in finding security flaws that it has decided not to launch it publicly. Claude Mythos Preview, an experimental version still in testing, demonstrated exceptional ability to identify critical vulnerabilities in popular software — which would represent both advancement and significant risk to digital security.

According to the company itself, the model would be "better than anyone except the most skilled humans" when discovering and pointing security breaches in browsers and operating systems widely used in the market.

AI technical capabilities in cybersecurity

Claude Mythos Preview is a general purpose language model, trained with a proprietary combination of public Internet information, public and private data sets, as well as synthetic data generated by other AI models.

According to Anthropic's documentation, the model represents a significant leap in performance when compared to other versions of the company, such as Claude Opus 4.6. His skills stand out especially in three areas:

But it was precisely code expertise that lit up warning signals internally.

AI identifies thousands of day-zero vulnerabilities

Anthropic's technical documentation reveals that Mythos has been able to identify "thousands of high-risk vulnerabilities" and zero-day failures (zero-day) — security breaches unknown by developers and therefore uncorrected.

Previous versions of Claude had already demonstrated this ability on a smaller scale, finding flaws in the Firefox browser and open source libraries. The vulnerabilities discovered were duly reported to the companies responsible and subsequently corrected.

The problem arises when this same technology could be used by malicious agents. Cybercriminals would have in hand a tool capable not only of identifying faults, but also of helping to create sophisticated ways to exploit them — or, when combined with autonomous AI agents, even performing steps of cyber invasions.

Glasswing Project: cybersecurity AI as a defense

Recognizing the dangers, Anthropic chose not to make Claude Mythos Preview available to the public. Instead, the company started the Glasswing Project, an initiative focused on using the capabilities of the model exclusively for security purposes.

Although complete details of the project have not been disclosed by the company, the approach represents a change in strategy: to transform a potential attack vector into a defensive tool, keeping the model under strict control and using it to identify and correct vulnerabilities before malicious agents discover them.

Debate on liability in the development of AI

Anthropic's decision rekindles discussions on ethical limits in the development of artificial intelligence systems. To what extent should companies advance in capacities that, although technically impressive, pose concrete security risks?

The Claude Mythos case illustrates a growing dilemma in the sector: increasingly powerful models require proportionally strict security protocols. Anthropic's transparency by documenting both the capabilities and the risks of the model sets an important precedent for industry.

The approach through the Glasswing Project suggests that the future of security AI may be not in the unrestricted release of powerful tools, but in its controlled and targeted use for collective protection.