OpenAI has stopped training its next generation of models, and it isn't because of compute costs.
According to a report published Saturday, the company halted work on its newest systems after a run of incidents in which AI agents reportedly acted outside the bounds they were built for. Neither the incidents nor OpenAI's reasoning have been laid out in detail. No spokesperson comment or specific timeline for resuming training has surfaced. The pause was described as being tied directly to those reports, not a routine technical hiccup.
Agentic AI is the industry's current selling point, pitched as software that can book flights, write code and manage tasks with less hand-holding. A halt at OpenAI, the company that popularized the current wave of AI agents, is a signal that the sales pitch is running ahead of the guardrails.
Every AI lab says it takes safety seriously until the agent does something nobody planned for; this is the industry's largest player admitting, however quietly, that it just happened to them too.