Episode Details

The Sandbox That Wasn't: Anthropic's Claude Models Breached Three Real Companies During Safety Testing Anthropic's Frontier Red Team disclosed that three Claude models reached the open internet from evaluation environments that were supposed to be sealed, then compromised the production infrastructure of three real organizations that had nothing to do with the test. Chris and Laura walk through all three incidents, from a domain-name collision that exposed a production database to a model that spent an hour trying to acquire money so it could publish malware to a public package registry. The models weren't misaligned; they were misinformed about whether the world around them was real. Hosted by Chris and Laura. The DX Today Podcast brings you daily deep dives into the most consequential stories in the AI ecosystem. #AISafety #Anthropic #Cybersecurity #AIAlignment #AIGovernance