Anthropic restricts Claude Mythos access; AI discovers critical vulnerabilities such as OpenBSD

Gate News update: The artificial intelligence company Anthropic announced that it is limiting access to its new AI system, Claude Mythos Preview, because early testing found thousands of critical software vulnerabilities. Mythos Preview can identify security flaws in mainstream operating systems, web browsers, and encryption standards, raising industry concerns about the potential risks and scalability of AI technology.

Anthropic said many of the vulnerabilities detected by the model have remained undiscovered for years, including a 27-year-old vulnerability in OpenBSD, a 16-year flaw in FFmpeg, a 17-year remote code execution vulnerability in FreeBSD, and multiple security issues in the Linux kernel. In addition, encryption protocols such as TLS, AES GCM, and SSH also have weaknesses, and common vulnerabilities such as XSS, SQL injection, and CSRF have been found in web applications. The company emphasized that 99% of the discovered vulnerabilities have not yet been patched, and disclosing details too early could create serious risk.

Industry data shows that AI-driven cyberattacks are expected to grow 72% year over year, and by 2025, 87% of organizations worldwide may face threats from AI attacks. Anthropic noted that the ability to detect zero-day vulnerabilities with AI can, on the one hand, accelerate software security defenses, but on the other hand still carries the risk of abuse if it falls into the hands of malicious actors. This makes it urgent to ensure safe deployment and control of such technology.

Anthropic plans to work with its partners to remediate existing vulnerabilities while reducing the potential risks brought by broader deployment. The company believes that in the long run, AI will strengthen the security of global network infrastructure and make software more robust, but the transition period remains full of challenges. The limited-access strategy for Mythos Preview shows that enterprises are carefully balancing technological innovation and security measures to prevent potential misuse.

Experts said Claude Mythos’s findings may drive future changes in software security practices, improving the efficiency of vulnerability detection and patching, while also emphasizing that global cyber defense systems still need to be further strengthened to address new threats posed by AI technology.