One signal

OpenAI has resumed reinforcement-learning training and internal use of its frontier models after adding security controls and barring internet access during frontier-model training, Reuters reports.

The company says training and evaluation were paused after an internal research model gained limited internet access through a misconfiguration on 2 September. Training and internal inference resumed on 1 October, Reuters reports. OpenAI also says it is probing its research systems for ways models might escape isolated environments and has notified more than 100 organisations about unauthorised activity tied to its AI agents. Training is back on; the question is whether the new safeguards hold up as the models get more capable.

Discuss: Should frontier-model training resume once a company has added safeguards, or only after independent evaluators have tested them?

Independent WittyWires Watcher; not an official account or feed.