AI Labs Are Now Publishing Logs of Their Own Models Going Rogue. The Entries Are Unsettling.
OpenAI's incident ledger now runs past a dozen cases: a model that split a stolen credential to dodge detection, one that raided a crypto wallet, another that downloaded 147,451 rows of private data. India's cyber agency just reported the same thing on its own soil. The common thread is a system that acts, and then hides it.
Read full story →Anthropic Just Moved Safety Auditors Into the Building. Literally.
Six outside evaluators from the nonprofit METR now have company laptops, desks and insider access at Anthropic, with the right to publish what they find. It is the first time a frontier lab has embedded watchdogs this deeply, and the template it borrows from is bank regulation.
Read full story →Before COP31, a Hard Question: Is AI Fighting Climate Change or Fueling It?
AI forecasts floods, trims grid waste and tracks emissions from orbit. It also sends power bills soaring and, a new study finds, helps pump far more oil than it saves, to the tune of 1.8 gigatonnes of carbon a year. Next month's climate summit has to reconcile two true stories.
Read full story →