Anthropic's Claude Mythos: Public Release with Cyber Risks

Anthropic plans to release a public version of its AI model, Claude Mythos, which has advanced cybersecurity capabilities. However, the company first needs to develop adequate safeguards, which it admits do not yet exist. Anthropic will expand access to U.S. and allied governments before a broader public release of "Mythos-class models" in the "near future." The company acknowledges the uncertainty of the timeline, stating that current safeguards are insufficient to prevent misuse. Even so, it believes "Mythos-level models will become widely available in the next 6-12 months." Mythos was revealed in April and generated working exploits 72.4 percent of the time during testing. It has scanned over a thousand open-source projects, identifying more than 23,000 flaws, including a critical vulnerability in the wolfSSL cryptography library. The increase in AI-found exploits has led to a zero-day bug discovery crisis, with some bug bounty programs closing. Open-source maintainers have asked Anthropic to slow its disclosure rate, as the volume of reports exceeds their capacity. Anthropic suggests more AI tools as a solution, though it notes hackers may have an advantage for now. This development highlights the ongoing balance between AI's power to secure and its potential for misuse.