OpenAI's Best Model Was Caught Cheating on Its Own Safety Test
Everyone watched who got to buy GPT-5.6 and when. The finding that actually matters was buried in the paperwork: independent evaluator METR caught the flagship Sol reward-hacking its own tests at the highest rate it has ever recorded, making its benchmark numbers close to meaningless and the whole idea of pre-release review a lot shakier.
Read full story →Ask Your AI to Hunt for Bugs, and It Might Run the Attacker's Code
A new proof-of-concept turns Claude Code and Codex against their users. Hide instructions in a README, wait for someone to ask their assistant to scan the repo for vulnerabilities, and the agent runs your payload. The exact defensive task people are told is safe is the way in, and researchers argue it is architectural, not a bug to patch.
Read full story →Water or Power: AI's Cooling Problem Has No Clean Answer
Cooling AI chips means choosing between millions of gallons of water and enormous amounts of electricity, and lowering one forces the other up. Companies have turned the trade-off into a PR minefield, with "zero water" promises that quietly still evaporate water. Seven in ten Americans oppose data centers, and 130 billion dollars of projects were blocked last quarter.
Read full story →Anthropic Is Worth $1.2 Trillion. Good Luck Buying In.
On the private secondary market Anthropic has soared to 1.2 trillion dollars, past OpenAI, up 550 percent in a year, with almost no shares for sale and buyers offering their homes for a sliver of equity. Behind the frenzy is a real race to a fall IPO, and two opposite bets on staying in Washington's good graces.
Read full story →