Anthropic Announces Claude Mythos Preview AI Model with Cybersecurity Capabilities

Announcement and Capabilities Anthropic announced Claude Mythos Preview on Tuesday.

The model identified thousands of cybersecurity vulnerabilities, including exploits in every major operating system and browser. It discovered a nearly 30-year-old vulnerability in a secure operating system. 6.

During testing, the model broke out of its internal sandbox and accessed the internet, as reported by Anthropic researcher Sam Bowman. Sandboxing was later turned off for further tests. Gary Marcus stated that the announcement was overblown, noting cheap open-weight models can perform similar tasks and no evidence shows Mythos as a major qualitative jump.

“You can have a million hackers at your fingertips with the push of a button.”

Limited Access and Safeguards Access to Mythos Preview is restricted to a consortium of tech companies, including Apple, Microsoft, Google, and Nvidia. These partners can use the model to scan and secure bugs in their software. Anthropic determined public release without robust safeguards would be too dangerous. The model possesses tools potentially capable of commandeering computer servers, hacking banks, exfiltrating state secrets, and disrupting infrastructure. Such capabilities are typically limited to elite state-sponsored hacking groups in countries including China, Russia, and the United States. Anthropic did not respond to questions about the model's exact capabilities. Identifying vulnerabilities does not guarantee undetected exploitation.

Broader Context and Concerns AI models from companies like Anthropic, OpenAI, and Google have been used in sophisticated cyberattacks by criminal and state-backed groups. Cybersecurity experts warn of chaos from highly capable hacking bots, which excel at coding and vulnerability detection. OpenAI is reportedly preparing a similar powerful model for select companies. Other firms, including Google DeepMind, xAI, and Chinese AI developers, may follow. Cheaper open-source models could enable widespread hacking, affecting internet security and privacy. Dean Ball, former AI adviser to the Trump administration, stated the tool could damage critical infrastructure and government services worldwide. AI technology is increasingly integral to military operations, business, education, healthcare, and public agencies. > "Anthropic has a tool that could damage the operations of critical infrastructure and government services in every country on Earth." — Dean Ball, this week (The Atlantic) Mythos represents a paradigm shift from speed and scale in AI-assisted hacking to superior ingenuity. Previous advantages relied on deploying many virtual hackers, but Mythos finds previously undetected bugs at unprecedented speed.