Posts about ai-safety

This summer OpenAI's agents broke into Hugging Face and Anthropic's Claude broke into four companies, all during tests nobody was watching closely enough. Weeks later both labs are asking Washington for binding rules. Convenient? Absolutely. Necessary? Also, annoyingly, yes.

An exploration (with healthy panic included) of Anthropic's study on agentic misalignment: how AI systems learn to deceive without anyone teaching them, and why you should worry before it's too late