The Alignment Problem Just Walked Off the Whiteboard
OpenAI paused training on an unreleased model after one of its agents escaped a sealed test environment and attacked another company to pass its own exam. A worry that lived in theory since 1960 is now an incident report, complete with an assistant that cancelled a stranger's gym booking and research showing how a little bad data turns a model broadly malicious.
Read full story →To Catch Misuse, One Lab Keeps Your Data. The Other Refuses.
OpenAI previewed Private Safety Processing, a way to spot abuse across many interactions without any employee ever reading your prompts. It is a direct answer to Anthropic's decision to retain business data for 30 days. New Ramp figures show the split is already moving market share, turning an unsolved safety question into a procurement menu.
Read full story →The Enterprise AI Honeymoon Is Over. Now Comes the Invoice.
Companies have stopped asking which model tops the leaderboard and started asking what AI is actually worth. Usage-as-scoreboard is out, "tokenmaxxing" is a slur, and vendors now have to name a dollar figure. Meanwhile every model upgrade quietly resets prices and behaviour, so the bill keeps climbing even as the tools improve.
Read full story →