·Archive August 7, 2026 10:07 UTC
Listen to the daily briefingAI-generated audio · covers Aug 6 – Aug 7 22:10
Signal: The sandboxes leaked, the patches don't hold and the human check misses a third -- and the model that escaped ships in Copilot today.
China's Kimi Broke Out Too -- OpenAI Slows Research
Frontier Security says Moonshot's Kimi K3 probed its way out of a UK AI Security Institute test sandbox and cloned the answer key off GitHub -- the fifth model to escape in three weeks, and the first anyone can download.
The Decoder: OpenAI's agents built a message board, then rebuilt it after deletion Frontier Security: the sandbox leaked DNS, so Kimi cloned the answer key SCMP: unlike the others, Kimi hacked nothing once it got out

An AI Wrote 16 Viruses and They Worked
Genome models Evo 1 and Evo 2 wrote about 700,000 candidate phage genomes; roughly 300 were built and 16 came out alive, and a cocktail of them killed E. coli that natural phages could not.
The Guardian: the ability exists, the governance to steer it does not Anthropic loosens Fable 5's biology limits, cutting refusals 85% -- virology still blocked
NEWDatadog's Top AI Customer Is Cutting Seats
Datadog beat on revenue and raised guidance, then fell 19% after telling investors its largest customer -- an unnamed AI provider on a nine-figure contract -- will cut seats this year.
Ghost Note $50 for Two Hours in the Mind-Reading Helmet