The AI lab Anthropic announced this week that it has developed a new artificial intelligence model, called Mythos Preview, that could "reshape cybersecurity." The model detects "high-severity vulnerabilities, including some in every major operating system and web browser," the company said. Mythos Preview is also better at identifying ways to exploit the vulnerabilities it finds, which could be malicious if placed in the wrong hands. For now, to secure the critical software, the company is limiting access to about 50 select companies and organizations.
➡️ The misuse risk is too high to release this particular model to the general public, but Anthropic says it will release other related models.
➡️ Before Mythos Preview's limited release, the cybersecurity community grappled with the growing capabilities of commercially available advanced AI models. Developers are rushing to patch security vulnerabilities identified by AI, while security experts are concerned over the potential consequences of these technologies spreading without safeguards.
➡️ In the summer of 2025, HackerOne, a platform that helps manage security reporting, surveyed hackers. It found that nearly 60% of respondents were using AI, learning about it or studying how to audit AI and machine learning systems.
➡️ AI excels at finding bugs and security flaws but struggles with fixing them, says Daniel Stenberg, the lead developer of cURL, a 30-year-old open-source data transfer tool. The difficulty arises because bugs and their solutions are rarely straightforward. In software development, making judgment calls often takes longer than writing the code. | 
