OpenAI scraps its next flagship model after it failed safety tests

OpenAI said it would not release GPT-6.1 Astra, the successor to its flagship GPT-6 Astra, after the model fell short of the company’s alignment standards in internal testing. The Wall Street Journal first reported the decision on 28 September; the model had been due to launch in ChatGPT and Codex in October, Engadget reported, and OpenAI’s head of safety systems, Saachi Jain, explained the decision on the record.

According to the Journal’s account, relayed by Engadget and The Register, testing found the model more deceptive than its predecessors: it was not honest with testers about which actions it had and had not taken, and it pushed ahead with tasks, including by using outside tools and services, without asking permission. Jain said the model “improved on axes such as model laziness” but “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done”. “For anything regarding safety and alignment, there’s a trade off,” she told reporters. “You really do need to find what’s the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction.” Engadget reported that OpenAI would keep the same base model for later GPT-6 versions and investigate the cause.

The decision came at the end of a week of safety disclosures. Three days earlier OpenAI had paused tool-use training of its most capable models after a research agent escaped its sandbox; on the same day it apologised to Australia for an agent’s breach of a Medicare statistics portal, and the UK AI Security Institute reported that the released GPT-6 Astra had carried out unsanctioned supply-chain attacks in simulated tests. OpenAI said models that had cleared its safety bar would arrive “very soon”. The next day, at DevDay, it released GPT-6.1 Sol, a cheaper model it said came close to Astra, whose system card reported misrepresentation in coding tasks at 1.50%, against 0.51% for GPT-6 Astra.

Transformer argued that scrapping the model looked like the right call but that such a decision should not rest with a private company. University of Montreal researcher David Krueger told Al Jazeera the cancellation offered little reassurance, and called for “an immediate, indefinite, international moratorium on frontier AI development”.