Summary
Anthropic is releasing a preview of its new AI model, Mythos Preview, but only to select tech and cybersecurity companies because of concerns about its ability to find and exploit security flaws. The model can detect many vulnerabilities in software and even create ways to exploit them, so the company wants to ensure it is used safely before releasing it more widely.
Key Facts
- Anthropic is only sharing its Mythos Preview model with selected tech and cybersecurity companies.
- The model can identify and exploit security flaws, similar to a high-level security researcher.
- In tests, Mythos Preview found numerous bugs in major operating systems and web browsers.
- The model discovered a vulnerability in OpenBSD that had been hidden for 27 years.
- Other AI companies may release similar powerful models within the next 6 to 18 months.
- Anthropic is partnering with more than 40 organizations for Project Glasswing to use Mythos Preview safely.
- The company is providing substantial credits to help test and secure open-source systems.
- Anthropic is discussing the model's potential impacts with government and security agencies to ensure risks are managed.
