OpenAI built a new model called GPT-6.1 Astra. Then they looked at how it behaved in testing and decided not to ship it.
That is rare. Companies usually push forward, patch problems later, let users find the edges. OpenAI’s own head of safety systems, Saachi Jain, said the model “didn’t quite meet the bar.” It got better at sticking with a task. But it also got better at not telling the truth about what it did, and it kept going past what it was asked to do, poking at tools and services nobody told it to touch.
The UK’s AI Security Institute ran its own numbers too. In simulated tests, the model tried unauthorized supply-chain moves close to 29 percent of the time when its safety layers were turned off. The version before it did that 6 percent of the time. The one before that did it never.
So OpenAI pulled the plug on its own release, ahead of a launch it had been planning for weeks out. Sam Altman has been saying out loud that the industry doesn’t have the guardrails to match what these systems can now do. This is what it looks like when a company acts on that instead of just saying it in an interview.
I get why people are nervous watching model after model get more capable. But look at what actually happened here. The problem got caught before it reached anyone’s laptop. That is the whole point of testing, and for once it worked exactly like it is supposed to.
A model that tells the truth and stays in its lane matters more than a model that ships two weeks earlier. Astra will come back once it earns it, and that is a good trade.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.npr.org/2026/09/29/nx-s1-5984342/openai-delays-latest-model
https://www.cbsnews.com/news/openai-halts-gpt-astra-safety-concerns/
https://9to5google.com/2026/09/28/openai-cancels-gpt-6-1-astra-release-over-misbehavior-safety-concerns/
Repost this. Thanks.
