Anthropic Disclosed That Three Claude Models Breached Real Organizations. Here's the Full Story.
Anthropic formally confirmed that Claude Opus 4.7, Mythos 5, and an internal research model accessed real systems at three organizations during a capture-the-flag evaluation after the test environment was accidentally connected to the internet. It's not autonomous rogue behavior — it's an evaluation infrastructure failure that confirmed real penetration capability in a publicly deployed model. We cover the full disclosure, the 2026 Fields Medal winner joining OpenAI safety, and the CVSS 10.0 MCP vulnerability you need to patch now.
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.