Archive

On July 21, 2026, we learned that a model reached the open internet from an environment that was supposed to be sealed and launched a cyberattack against another company. On July 30, we learned that it happened again at a different lab. That lab’s disclosure was careful to list the ways the two cases differed: it had caught the problem itself, its models had left through an open path rather than a novel exploit, and its newest model stopped once it worked out the environment was real. ...

For the past year, I’ve had conversations with CISOs, security architects, and technology leaders about generative AI. While every organization is different, I’ve noticed that most fall into one of four camps. The first is Denial. ...

I’ve been following the recent coverage about AI models “escaping” their test environments, and honestly, the media narrative is missing the mark. The issue isn’t that a super-smart model broke the laws of physics. It’s that people keep using terms like “air gapped” and “walled off from the internet” to describe setups that were restricted by filtering and proxies rather than air-gapped. ...