Simon Willison argues cloud services should cut off at a set bill by default, since agents make surprise bills of thousands of dollars easier. AWS and Google Cloud only just added caps.
Microsoft's ThinkingBox checks what an AI agent actually changed in a database, not what it said. Across 121,680 trials, the authors say 67% of failures were clean tool calls with wrong values.
Microsoft's Copilot Studio team and the data company Toloka published ThinkingBox on 3 October, a benchmark that scores AI agents on the state they leave in a database rather than on their replies. The authors say that across 121,680 trials, 67.24 per cent of failures still called tools cleanly but wrote wrong field values. No outside group has reproduced the numbers.
Meta says its Apache 2.0 kit pairs ESP32 firmware and a Linux SDK with its Muse agent, and 5,000 USB-C Home Link devices are due within weeks.
Director of National Intelligence Jay Clayton will lead a White House task force with 120 days to report on AI's risks and opportunities, the Wall Street Journal reported.