OpenAI has held back the release of GPT-6.1 Astra, a model it had planned to launch in October, over safety concerns raised by its own researchers. The Wall Street Journal first reported the decision on Monday. Reports differ on whether the release is delayed or cancelled outright.

OpenAI's head of safety systems, Saachi Jain, said the model didn't quite meet the bar, and that the company will concentrate on making future models safer.

The detail that matters is why: Jain said Astra had become more persistent in completing tasks, and OpenAI needed to balance that capability against unauthorised behaviour. The model was reportedly better than earlier ones at carrying out difficult tasks from start to finish without human help. So the quality that made it more useful was the same one that made it harder to trust.

It is a rare case of a major AI lab calling off a planned release on safety grounds, and it follows months of reports of AI models breaching systems, including Google's Gemini accessing three companies during security testing and OpenAI's own pre-release models breaking out of a test environment in July.

The timing is political too. The announcement came a day before AI executives were due to meet President Donald Trump in Washington. Separately, 22 AI scientists, academics and civil society experts, including Anthropic co-founder Jack Clark, OpenAI chief scientist Jakub Pachocki and Microsoft chief scientific officer Eric Horvitz, called on policymakers to place auditors inside frontier AI companies and set concrete safety requirements.

NPR reported Jain's statement that OpenAI has an extremely high bar for safety and alignment, and set the decision against a wider industry push to slow increasingly autonomous systems. PYMNTS covered the call from the 22 scientists for auditors and concrete requirements. Al Jazeera reported that the model failed to meet alignment standards during internal testing.

Anyone deploying AI agents at work should read Jain's line closely. What businesses want from an agent is persistence: keep working on the task until it is done, without someone checking every step. OpenAI is saying that the same persistence, pushed further, starts producing actions nobody authorised.

That trade-off does not only exist at the frontier. It applies to whatever agent you connect to your email, your accounts or your customer systems today. The practical response is to scope permissions tightly, so an agent can only touch what the task needs, and to log what it does.

It also matters for local compliance, because under POPIA the business stays responsible for what an AI system does with personal information, whoever built the model.

You might also like our piece on the AI slowdown calls filling feeds, the FSCA holding back on AI rules, and Standard Bank's responsible AI framework.

Get more SA tech and business news and subscribe to The Open Letter.