Anthropic has uncovered thousands of software vulnerabilities through its new AI model, Claude Mythos, which will not be publicly released. This precaution comes after the model revealed flaws that had remained undetected for years, some dating back 27 years. In response, the San Francisco-based startup is collaborating with cybersecurity firms to enhance defenses against potential hacking threats.
At a recent conference in San Francisco, Mike Krieger from Anthropic Labs emphasized the model's defensive potential, stating it allows cybersecurity specialists to utilize its capabilities proactively. The vulnerabilities identified include subtle issues that traditional testing methods might overlook; for instance, Mythos detected a flaw in video software that had undergone over 5 million tests without being recognized by its developers.
As part of the initiative, Anthropic has partnered with companies like CrowdStrike, Palo Alto Networks, Amazon, and Apple, among others, in a project named Glasswing. The collaboration also involves Cisco and Broadcom, along with the Linux Foundation, highlighting the urgency and significance of addressing these cybersecurity challenges collectively.