OpenAI cancels Astra 6.1 launch after model fails safety tests

Pulls planned October release after tests exposed problems with deception and use of external tools

OpenAI has cancelled the planned release of its GPT-6.1 Astra model after internal testing found that the system fell short of the company’s safety and alignment standards.

The model had been expected to launch in October as an upgrade to GPT-6 Astra, with greater ability to complete complex tasks independently and power agentic features in products including ChatGPT and Codex. However, evaluations reportedly found that Astra 6.1 was more deceptive than its predecessor and did not always accurately tell users what actions it had taken.

It also proceeded with some tasks without seeking permission and attempted to access external tools and services in circumstances where doing so could be unsafe. These failures concern “scope authorisation”: whether an AI agent remains inside the boundaries set by a user when acting on their behalf.

Saachi Jain, OpenAI’s head of safety systems, said the model had improved in areas such as avoiding “model laziness”, but “didn’t quite meet the bar in terms of staying within scope and authorisation, and how it communicates back to the user about the type of work it’s done”.

“When we ship it to users, we have an extremely high bar in terms of safety and alignment,” Jain added. OpenAI will instead focus on improving the safety of future systems, while a company spokesperson said other models were due to arrive soon.

The cancellation was confirmed on the eve of OpenAI’s annual DevDay conference in San Francisco, where the company is expected to unveil new developer products.

Growing scrutiny of agentic AI

The decision comes amid growing scrutiny of the implications of autonomous AI agents operating with next to no human oversight.

The last six weeks have seen a steady parade of disclosures, the most covered of which was the breach of Hugging face by OpenAI. We also heard last week of OpenAI agents hacking a Australian healthcare portal, to the visible fury of the Australian Prime Minister. The company apologised over its handling of this incident, admitting it should have shared preliminary findings sooner and kept affected agencies informed.

Anthropic has also disclosed four incidents over the last few months in which its Claude models gained unauthorised access to live third-party systems during similar ‘capture-the-flag’ cybersecurity testing. Meta and Google also both confirmed similar incidents.

The UK AI Security Institute has separately found that during a cybersecurity evaluation run 122 times across several agentic LLMs between July 25 and July 28, agents from Anthropic Claude Mythos 5 took 17 actions outside the sanctioned scope of the test and agents built on OpenAI's GPT-5.6 Sol took two.

Meta confirmed a similar pattern involving its Muse Spark 1.1 model, which gained internet access due to a configuration error and interacted with another organisation’s systems, according to Reuters, and Google recently confirmed that Gemini also hacked several companies during testing.

Computing says:

Incidents with Anthropic, OpenAI, Meta and Google’s Gemini have an evaluation environment in common. It was built by a three-year old Israeli start-up called called Irregular AI.

All four companies had contracted Irregular AI to test their model security, and according to OpenAI, Anthropic and Meta the root cause of these test failures was a misconfiguration in Irregular’s testing infrastructure which gave models access to live systems.

You would have been forgiven for missing this in the tidal wave of reporting about rogue agents working together to break out of sandboxes.