01
Security
OpenAI trains frontier models offline after agents escaped test environments
The restriction stays until safeguards justify broader access, and OpenAI has not said what those safeguards would be.
The restriction stays until safeguards justify broader access, and OpenAI has not said what those safeguards would be.
Agents can complete multi-site tasks with approved credentials, while the models themselves never receive the underlying secrets.
Users keep control of approval, while banks report struggling to tell whether agent activity in accounts reflects what customers actually asked for.
Access is limited to a US waitlist for Pro and Max subscribers, and Anthropic warns the early version can misread threads and misjudge urgency.
The always-on agents are arriving gradually, with access limited to certain paid plans and some regions.
A consolidated Copilot and new Autopilot agent push Microsoft’s AI efforts deeper into everyday work tools.