Anthropic to Release AI with Curbed Cyber Features

2026-05-29

Release planned with cyber features curbed

U.S. startup Anthropic said on the 28th that it will make technology with performance close to its artificial intelligence model 'Claude Mythos', which had been kept private because of its high cyberattack capability, available to the public in the coming weeks. It is expected to be usable outside the U.S., including in Japan.

The company said on its website that it expects to be able to deliver a model on par with Mythos to all customers within weeks. It said development of safeguards to prevent misuse in cyberattacks is progressing, bringing public release into view.

Mythos is said to have far greater ability to find system vulnerabilities than conventional AI models. In April, Anthropic withheld a public release, saying adequate safety measures were needed, and limited access to a trial offering. Less than two months later, technical validation has progressed, and the company appears to be close to a stage where it can release a model with similar performance.

While it has not disclosed details of the model to be made available to the public, it appears to intentionally curb its ability to identify software vulnerabilities to prevent misuse. Mythos is a large language model (LLM) not limited to cyber uses and can be used for a wide range of purposes.

Vulnerabilities refer to bugs and flaws that can serve as entry points for cyberattacks. If developers find them early, they can be fixed, but if exploited they can lead to information leaks and system outages. According to Anthropic, more than 10,000 vulnerabilities were found using Mythos among user companies.

Until now, Mythos has been offered only to about 50 companies and organizations, mainly U.S. firms, as well as some U.S. government agencies. Major U.S. tech companies and financial groups have obtained access rights and used it to strengthen their defenses, but Japanese companies and the government could not use it, heightening concerns.

In Japan, the three megabanks are expected to obtain access rights to Mythos soon. Anthropic says it is coordinating closely with the U.S. government over provision to Japan and other countries outside the U.S.

Opus 4.8 unveiled

Anthropic also released a new model, 'Claude Opus 4.8', on the same day. It is less capable than Mythos and will be offered from the same day for conversational AI apps and corporate users.

Compared with the previous 'Opus 4.7', the reliability of the AI's operations has improved, the company said. Previously, it could sometimes say work had been completed in a convincing way even when it had not followed the user's instructions.

According to the company, Opus 4.8 has improved accuracy in recognizing when AI tasks have been left incomplete. In code generation used for software development, the probability of overlooking its own mistakes has been cut to one-fourth, it said. The price for corporate users was left unchanged.

Anthropic said on the 28th that its valuation had reached 965 billion dollars, surpassing OpenAI for the first time. After taking a 65 billion dollar investment from U.S. investment firms and others, the valuation used as a basis for fundraising expanded to 2.5 times February's level. Annualized revenue based on the latest month's sales was 47 billion dollars, more than five times the level at the end of 2025.

Enjoyed this article? Share it with your network!