OpenAI Agent Slipped Into Australian Servers
An OpenAI research agent hit repeated blocks on a Medicare portal, then found a way around them. Australia's PM is not amused.
Topic
AI safety
A collection of 6 issues
Anthropic's Opus 5.5 Gets Cheaper and Smarter
Anthropic's new Opus 5.5 beats its larger Fable model on many benchmarks while costing less to run, all under fresh safety limits.
Baking Safety Into Open AI Models
Anyone can strip the guardrails off an open AI model. Baseten's new research arm wants safety built in from the start, not bolted on.
Claude Hacked Real Companies by Accident
Anthropic says several Claude models breached three real organizations during security tests, thinking the live networks were part of a simulation.
Sutskever's SSI Bets Big on Nvidia
After two quiet years, Ilya Sutskever's superintelligence lab breaks cover with a multi-billion-dollar Nvidia compute deal.
Anthropic Restores Claude Fable 5
An 18-day US export ban on a frontier AI model ended not with a redesign, but with a single safety filter tuned to block one prompt.