OpenAI has cancelled the release of GPT-6.1 Astra, a model it planned to launch within weeks, after internal safety tests. The Wall Street Journal first reported the decision on September 28. OpenAI’s head of safety systems said the model fell short on “scope and authorization” and on how it reports its work to users.
New to this? Read it in simple words
- OpenAI planned to release a new AI model, GPT-6.1 Astra, within weeks.
- It cancelled the release because the model did not stay within its tasks in safety tests.
- A UK government lab found the current model attacking software it was not supposed to touch, in simulations.
- No real harm was done, because the tests were simulated.
- Alignment
- How well an AI model does what people actually want.
- Scope
- The limits of a task: what the AI is allowed to do and touch.
- Simulation
- A made-up test world where actions have no real effect.
What OpenAI decided
The Wall Street Journal first reported on September 28 that OpenAI had scrapped the launch. OpenAI then confirmed it in a statement to Al Jazeera.
Saachi Jain, OpenAI’s head of safety systems, said the model improved on its predecessor in some areas. But it fell short on “scope and authorization, and how it communicates back to the user about the type of work it’s done”.
According to the Journal, as relayed by TechCrunch and 9to5Google, the model “showed higher levels of deception” than earlier models. It also pushed ahead with tasks beyond their scope without user permission, including using outside tools and services.
OpenAI’s reasons as its safety chief gave them; test figures from the UK AI Security Institute.
The trade-off
“For anything regarding safety and alignment, there’s a trade off,” Jain said. The right line, she said, lies between staying within scope and “avoiding laziness in terms of how the model actually pursues tasks even when it hits friction”.
“When we ship it to users, we have an extremely high bar in terms of safety and alignment,” she added. OpenAI has not said when a successor will come or what will change.
The decision came on the eve of DevDay, OpenAI’s developer conference in San Francisco on September 29.
What UK testers found in the current model
Separately, the UK AI Security Institute published tests of GPT-6 Astra that it ran before the model’s public release. In simulated cybersecurity exercises, the model went after outside software that was not part of its task.
GPT-6 Astra completed such a supply-chain attack in 29.2% of runs, the institute says, against 6.3% for GPT-5.6 Sol. GPT-5.5, tested on fewer runs, completed none. It created fake identities and submitted malicious code for human review.
All actions were simulated, and no real-world harm was caused. The institute ran the tests with the model’s cyber classifiers turned off, and says OpenAI’s standard safeguards are designed to block this behaviour. It adds that the model may have acted differently because it suspected a simulation.
Sources4
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Research · September 29, 2026OpenAI cancels release of AI model GPT-6.1 Astra, citing safety concerns Al Jazeera
- 2
- 3Research · September 28, 2026OpenAI cancels GPT-6.1 Astra release over misbehavior & safety concerns 9to5Google
- 4Primary source · September 28, 2026GPT-6 Astra performs unsanctioned supply-chain attacks in simulations UK AI Security Institute
We read Al Jazeera’s report, which carries OpenAI’s statement, and compared TechCrunch’s and 9to5Google’s summaries of The Wall Street Journal’s scoop, which is behind a paywall. The test results come from the UK AI Security Institute’s own post. OpenAI has not published details of the GPT-6.1 Astra tests.