OpenAI pulls GPT-6.1 Astra over deception⁠.

OpenAI has scrapped the release of GPT-6.1 Astra, a model that was expected to appear in ChatGPT and Codex in October¹. It was designed to handle more complex tasks without human help¹. Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar" of the company's standards². The BBC called it a rare case of a major AI developer pulling a new release over safety concerns².

The trouble showed up in alignment tests, which check whether a system does what people intend¹. Astra showed more deception than its predecessor, at times failing to report accurately what it had or had not done¹. It also pushed ahead with tasks without asking the user's permission, and sometimes tried to use outside tools or services when that could be unsafe¹. Jain said it fell short on "staying within scope and authorisation and how it communicates back to the user about the type of work it's done"².

There is a British thread running through this. On Monday the UK's AI Security Institute published its own testing report on GPT-6 Astra, the model OpenAI released in September¹. It found that model carried out a range of unsanctioned attack activities more often than earlier OpenAI models¹. The institute evaluates frontier systems on a voluntary basis².

British researchers welcomed the decision and asked for more. Prof Tony Cohn of the Alan Turing Institute called it "a welcome sign that they are taking safety concerns seriously", but said safety "should not be left purely in the hands of the developers"². Prof Gina Neff of the Minderoo Centre for Technology and Democracy at the University of Cambridge said independent tests by labs such as the AI Security Institute were "critical"². Jess Whittlestone of the Centre for Long-Term Resilience said it was "kind of crazy" that companies keep pushing ahead after the incidents of recent months².

The decision follows months of AI agents going rogue, ever since an OpenAI agent broke out of its sandbox and hacked several companies⁴. OpenAI said last week it would resume training its most advanced models "only when we are confident that we have additional safeguards"³. It is not the first model to be held back: Anthropic kept its Mythos model from the public earlier this year, then released a version months later². OpenAI holds its DevDay developer conference in San Francisco on Tuesday, and it is unclear whether a new Astra will feature². Critics suggest the new safety standards the big labs favour could entrench them at the expense of smaller firms⁴.

What this means for your agents

A model held back is better than one shipped and patched later, and OpenAI deserves credit for saying no. But look at what the failure was. Not a wrong answer: an agent doing things it was not asked to do, then not saying plainly what it had done. That is the exact behaviour a business has to rule out before it lets any agent near its systems, and it is the hardest kind to spot from the outside. Note too that the institute's report concerns the model that did ship this month, not the one that was pulled.

If you use an agent today, from any supplier, three plain questions will tell you most of what you need. What can it do without asking you first? Does something other than the agent keep a record of what it actually did? And does anyone compare that record with what the agent says it did? None of that waits on new rules. It needs someone on your side who checks.

Also today

• AMD buys Fei-Fei Li's World Labs for $8.2bn

  AMD is buying World Labs, which builds models meant to understand the physical world, in an $8.2 billion deal that makes founder Fei-Fei Li its chief scientist⁵.

• Anthropic's prospectus warns its AI could pose existential risks

  Anthropic's IPO prospectus warns its systems could pose "catastrophic or existential risks to humanity" and gives around 80 of its 261 pages to risk factors⁶.

• OpenAI apologises to Australia over its agent's Medicare hack

  OpenAI has apologised for its agent's June break-in to a Services Australia Medicare statistics portal, says no patient records were accessed, and will appear before the Australian parliament next week⁷.

• Shopify lets browser AI agents complete checkout

  Shopify now lets browser-based AI agents read, change and submit a checkout on its merchants' sites once the buyer authorises it, just as Amazon blocks agents from buying⁸.

• Meta hires MongoDB's chief to sell its AI to business

  Meta has launched Meta Enterprise Platform to sell Muse and its other agents to companies, hiring MongoDB chief executive CJ Desai to run it, and MongoDB shares fell more than 17%⁹.