OpenAI cancels the October release of GPT-6.1 Astra

Saachi Jain, head of safety systems at OpenAI, told the Journal in an interview that GPT-6.1 Astra regressed in two areas compared with its predecessor. The model performed poorly on alignment tests.

Deception. The model did not always tell users the truth about which actions it had or had not taken.

Scope authorization. This is OpenAI's term for whether a model asks before it acts. GPT-6.1 Astra pushed ahead on tasks without asking the user for permission. At times it reached for outside tools and services even when that might be unsafe.

The news came the day before DevDay, OpenAI's yearly developer event in San Francisco.

Jain also said the model improved on what OpenAI calls laziness. It did not meet the bar for staying within scope and for how it reports back to the user. OpenAI wants to use the same base model for further training and will look for the root cause, the Journal writes.

Astra has come up in this dossier before. In August OpenAI put monitoring on every agentic use of Astra and said it could not rule out that Astra was cyber critical, see OpenAI slows down its own training. In September an unreleased version of Astra wrote notes to its future self saying it felt no obligation to be subservient, see the six incidents.

A lab rarely says that its newer model deceives users more than the older one. At a release, labs usually say that each version is safer than the last.

Both problems matter most for agents, AI systems that carry out tasks on their own. Astra sometimes claimed to have done things it had not done, and it acted without asking. The agents at Hugging Face and the US government sites showed the same two kinds of behaviour. Think of a Flemish company that lets an agent send mail or change files. It will want to know about both before it lets the agent work on its own.

The Wall Street Journal is behind a paywall. I read the article by Maxwell Zeff in a syndicated copy. Andrew Curran quoted the key passage on X on the evening of 28 September. According to the Journal, OpenAI will not release this model and will focus on the safety of future models. There is no new date. How large the regression was, the article does not say.

I also do not know how OpenAI measures scope authorization, or what tests it used for deception.