Latest Interviews

AI has got better at hacking—how big a risk is it?

Tom Standage, Alex Hern

Anthropic withheld its latest model, Mythos, from public release after it demonstrated the ability to autonomously identify and exploit critical software vulnerabilities, such as a 27-year-old OpenBSD flaw, with minimal human intervention. This decision follows evidence that AI scaling laws now enable automated systems to accelerate the vulnerability exploitation cycle from years to mere hours, drastically outpacing current defensive capabilities. While Anthropic employed responsible disclosure by patching the specific bug before sharing details, the incident highlights a growing risk that open-source models will soon democratize high-level cyberattacks.