Anthropic's Mythos Preview sparks global cybersecurity alarm

In June 2024, a cyber-attack on a pathology services company caused chaos across London hospitals. More than 10,000 appointments were cancelled. Blood shortages followed and delays to blood tests led to a patient’s death. The incident underscored how digital smashups in the health sector can spill into the physical world, reminding the public that the safety of everyday care depends on resilient software. While this is not a new threat, a fresh wave of AI tools is shifting the balance between defense and offense.

Anthropic has unveiled Claude Mythos Preview, a powerful AI model the company says is too dangerous to release publicly because of its cybersecurity and cyber-attacking potential. The move signals a shift in how the tech industry handles highly capable AI: instead of a wide public release, access is limited to partners for defensive testing and threat research.

Advocates of Mythos say the system can autonomously scan software for vulnerabilities, spotting flaws across major browsers and operating systems, and even surfacing bugs in the Linux kernel. In one account, it reportedly identified a 27-year-old bug in critical security infrastructure, illustrating how such a tool could reveal weak points across the internet.

Security researchers warn that if this capability falls into the wrong hands, it could turbocharge cyberattacks, enabling amateurs to disrupt banking, streaming, transport, and other essential services. Experts have described the threshold as a turning point that changes how urgency is measured in cybersecurity.

Anthropic argues the risk is real enough to justify controlled access and protective measures. The company has launched a defensive program called Glasswing, aiming to fix flaws before attackers exploit them and to keep more dangerous features off public shelves. The debate taps into broader questions about AI governance, risk, and who gets to wield weapons that can rewrite system security.

As governments and firms weigh safeguards, the public is left facing a stark question: how do we harness powerful AI tools to improve security without giving criminals a shortcut to breach the systems we rely on every day? The answer is still evolving, and the conversation is likely to intensify in the months ahead.

#Anthropic # ClaudeMythos #CyberSecurity