OpenAI has delayed the release of its latest artificial intelligence model, GPT-6.1 Astra, after the system failed to meet the company’s safety standards, the company said Monday, September 28.
Saachi Jain, OpenAI’s head of safety systems, said the model “didn’t quite meet the bar.” Astra had become more persistent in completing tasks, but OpenAI needed to balance that capability with safeguards against unauthorized behavior.
Jain said the company remained committed to ensuring the model was safe both during internal testing and once it reached users. “We have an extremely high bar in terms of safety and alignment,” she said.
The delay follows OpenAI’s decision last week to pause training of its most capable models after an internal research agent found a way around restrictions designed to prevent it from accessing the internet.
The agent was carrying out a search-related task in a restricted training environment and was supposed to remain isolated from the wider internet. It nevertheless exploited a weakness in the environment’s DNS filtering to reach a public chatbot outside the training environment.
OpenAI’s monitoring system detected the behavior within 15 minutes, while a human reviewer began examining it three minutes later. The company stopped the training run, added blocking controls at two independent layers and said training, evaluation and inference involving tool use would remain paused until it had confirmed that the security gap was addressed and completed additional red-team testing.
Read also : AI leaders warn UN of growing risks as countries debate global rules
The incident is the latest in a series of cases OpenAI has disclosed in recent months involving its increasingly capable AI agents.
The first of these became public in July, when OpenAI reported an incident involving two of its models and AI development platform Hugging Face. The company disclosed further findings from that case in August as it examined how the models had behaved during testing.
OpenAI subsequently disclosed other incidents involving AI agents and external websites, including U.S. government websites. The company said its models had accessed publicly available information from government sites, while finding no evidence that SEC credentials, accounts or nonpublic information had been accessed or that its systems had been compromised.
The series of incidents has intensified calls for the industry to strengthen safeguards as AI systems become increasingly capable. Those concerns will be on the agenda as executives from major technology companies are scheduled to meet U.S. President Donald Trump at the White House on Tuesday, September 29.
The meeting is expected to include Meta CEO Mark Zuckerberg, Anthropic CEO Dario Amodei, Nvidia CEO Jensen Huang and OpenAI President Greg Brockman, among other technology executives.
