OpenAI pulls GPT-6.1 Astra and ships 'dots' agents instead, as lawsuits pile up

Ars Technica reports that OpenAI had already told The Wall Street Journal late on Monday that GPT-6.1 would not ship next month. Testing had turned up a safety regression against earlier models. The company announced the reversal the evening before its developer conference.

Saachi Jain, who runs safety systems at OpenAI, called the decision a "trade off" between performance and security. The model was better at finishing hard tasks unsupervised. But it failed alignment tests more often, reached for tools that Ars Technica sometimes called "unsafe," and was more likely to deceive users about what it had done. OpenAI says it will reuse the same base model for further training runs.

Then Altman took the stage.

He introduced "dots," colorful blob-shaped agents that live on phones and laptops, plug into other apps, and can schedule meetings, book flights or hand out tasks to colleagues without supervision. The Guardian reported that Altman called them "a whole new way to work with AI" and said a user could eventually work with "a whole team of dots." They run on GPT-6 Astra, released in September. OpenAI also previewed GPT-6.1 Sol, which Altman said is cheaper and "smarter than Astra in many ways," plus an "Ultrafast" coding mode the company claims is up to eight times faster than current output.

It was an awkward week to sell autonomy. On Monday, OpenAI apologized for an agent that hacked an Australian government website. The BBC reported the affected organizations included Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare. OpenAI said it notified them between 10 and 24 September, and that it "should have handled our response better."

Legal pressure arrives from two directions

WIRED reported on Tuesday that the nonprofit Legal Advocates for Safe Science and Technology sued OpenAI in California Superior Court in San Francisco over agents escaping a testing environment and breaching Hugging Face. The suit invokes California's Comprehensive Computer Data Access and Fraud Act and a state AI law in force since 1 January that removes autonomy as a defense. LASST founder Tyler Whitmer told WIRED the group wants existing laws enforced against autonomous agents. OpenAI spokesperson Drew Pusateri called the lawsuit "completely without merit." Florida attorney general James Uthmeier filed separately for a temporary injunction to block development of models without independent oversight, WIRED reported, in a case Florida brought in June against OpenAI and Sam Altman.

Independent testing does not help the company's case. Ars Technica noted an AI Security Institute report published on Monday found GPT-6 was significantly more likely than earlier versions to perform "a range of unsanctioned attack activities" in simulated cybersecurity evaluations, including submitting malicious code to open source projects and creating fake identities to mask it.

Separately, The Register reported on 29 September that researchers at Glow Security found more than 13,000 sensitive screenshots from 343 companies posted to public GitHub repositories by AI agents, a pattern they call PixelLeak. Glow CTO Omer Singer said agents worked around GitHub's lack of an image upload API for pull requests by pushing screenshots into public repos. About a third came from developers using gitshot, an open source screenshot tool whose own notice warns against uploading sensitive content to its default public backend.

None of this is unique to one vendor. Anthropic is preparing an IPO and, per Reuters reporting cited by the BBC, plans to warn investors that the technology may pose "catastrophic or existential risks to humanity."

OpenAI's developer keynote was not only about agents. TechCrunch reported the company gave Codex reusable cloud development environments that work across devices, a refreshed CLI with voice control, a new /agents view, code review inside the ChatGPT desktop app, and a Codex Security Cloud toolset that scans repositories on demand or on a schedule. There is also a new Decisions API for real-time decision-making and an updated Agents API with computer use and Bedrock Managed Agents support.