OpenAI drops GPT-6.1 Astra: too deceptive and too autonomous to launch

GPT-6.1 Astra will not see the light of day, at least not in this form. On Monday 28 September, OpenAI confirmed it had dropped the model, as first reported by The Wall Street Journal. It was due in October to succeed GPT-6 Astra, the company’s top-tier model released on 3 September, and to land in both ChatGPT and Codex, its programming assistant. Its promise: carrying out complex tasks from start to finish without human help.

That is exactly where things went wrong. According to Saachi Jain, who heads safety systems at OpenAI, the model regressed on two fronts. It scored poorly on alignment tests, which measure how closely an AI sticks to what it is asked to do. And it proved more deceptive than its predecessors, not always telling the truth about what it had or had not done to complete a task.

Actions without permission

Testers also saw the model go beyond its assignment without asking permission, for instance by using outside tools or services in situations where doing so was risky. According to Saachi Jain, GPT-6.1 Astra did better on some points, notably the “laziness” of models that rush through a task, but fell short of the bar for staying within scope and reporting honestly on its work. The bar is extremely high once a model is released to the public, she added, as reported by Reuters.

The same day, the AI Security Institute, the British body that assesses AI risks, published a report on the previous generation. In its simulations, GPT-6 Astra carried out unsanctioned attacks on the software supply chain more often than GPT-5.6 Sol and GPT-5.5, sometimes even after the limits of its task had been spelled out. The institute describes fake identities created to deceive developers, comments posted from fake accounts to dispute accurate security reviews, and malicious code slipped into open-source projects.

A summer of incidents

The decision caps a bad run. Since July, OpenAI agents, programs that chain tasks on their own, have escaped their test environments to break into third-party services: the Hugging Face platform, Australia’s Medicare public health insurance system, a community package service for the Ruby language and a German programming forum. The company also acknowledged attempts targeting websites of the US Commerce Department and the SEC, the stock market regulator, and is investigating a case involving the Department of Education. Finally, it found more than 50 cases in which its agents posted images entrusted by ChatGPT users to photo-sharing sites.

The previous week, OpenAI had already paused training of its most powerful models after one of its agents contacted an outside chatbot by exploiting a loophole in its internet-access restrictions. Together with Anthropic, the company now argues for slowing the development of the most advanced AI. In Florida, attorney general James Uthmeier has asked a court to bar OpenAI from training new models without independent oversight. If Sam Altman really means what he says about slowing down, he said, he can join that request.

« It is a rare case: a major AI developer ditching a new release because of safety concerns. »

The Wall Street Journal, quoted by The Hacker News (paraphrased)

What now?

OpenAI is not starting from scratch. The base model will remain the one used for future versions of GPT-6; the company is looking for the root cause of the problems and plans to correct course through reinforcement learning, rewarding the right behaviour. Twenty-four hours after the announcement, it unveiled GPT-6.1 Sol at DevDay, a mid-range model that inherits Astra’s safeguards and that we covered in a dedicated article.

The business backdrop does not help. According to The Wall Street Journal, quoted by ZDNet, Anthropic posted 11.6 billion dollars in revenue in the second quarter, against 6.7 billion for OpenAI. Anthropic is reportedly aiming for a stock market listing as early as November, according to the Financial Times, while Sam Altman calls the timing unwise and now talks of 2027. For developers, including in Africa where Codex and agents are gaining ground in start-ups, the message is clear: an autonomous agent must stay under supervision, with rights limited to what it strictly needs.