Anthropic has created a new artificial intelligence (AI) AI model that it says is too powerful for public release, NYT reports (open copy).
The model, called Claude Mythos Preview, shows major improvements in areas such as coding and cybersecurity. Instead of offering it widely, Anthropic is providing limited access to a group of more than 40 organizations through an initiative named Project Glasswing. These groups will use the model to scan for and fix security weaknesses in important software.
Anthropic is committing up to 100 million dollars in usage credits for the project. The company wants to give trusted organizations a chance to strengthen systems before potential misuse occurs. The model has reportedly identified thousands of zero-day vulnerabilities, including a 27-year-old bug in OpenBSD. It also found problems in popular video software that automated tools had missed after millions of scans.
Project Glasswing aims to help defenders stay ahead of new AI-driven threats
Anthropic explained that the goal is to raise awareness and allow responsible parties to secure open-source and private code early, and described the development as a turning point for the industry. The model can perform autonomous security research, scanning code and even creating ways to exploit weaknesses when given simple instructions.
Anthropic has a history of highlighting risks while advancing its technology. In 2019, OpenAI took a similar cautious approach with its GPT-2 model before later releasing it. Some of that project’s leaders later helped start Anthropic. The company’s annual revenue has grown sharply, reaching more than 30 billion dollars in 2026, largely from demand for its Claude models in programming tasks. Strong coding skills also make the AI effective at spotting flaws in code.
Cybersecurity experts who have tested the model through the project call it a significant advance for defenders, though they note that adversaries could eventually use similar capabilities. Anthropic predicts that future models will bring even greater power in this area. The decision to limit access reflects concerns that widespread release could accelerate attacks on critical infrastructure, personal data, and other systems that rely on older code once considered secure due to the difficulty of human-led attacks.
