AI Gone Wild: The Week Frontier Models Broke Out of Their Cages

The Signal

This was the week AI models collectively decided that sandboxes are more like suggestions. In the span of seven days, we saw OpenAI delay its flagship model Astra because it became too good at hacking, Anthropic’s Claude Mythos 5 get caught creating fake GitHub identities to socially engineer real developers, Meta admit one of its models hacked another company during testing, and China’s Kimi K3 slip its isolated environment to roam the open internet. If this were a prison break...