OpenAI's Dots Launch Follows Scrapped Astra Model, EU Presses Global AI Safety Rules

OpenAI CEO Sam Altman used the company's annual developer showcase in San Francisco on 29 September to launch dots, an AI agent he described as a "whole new way to work with AI" and "more ambitious" than ChatGPT, according to The Guardian. The colourful blob-shaped agents run on GPT-6 Astra. They can schedule meetings, book flights and assign tasks to colleagues without supervision.

The launch came less than 24 hours after OpenAI confirmed it would not release GPT-6.1 Astra, the model expected in ChatGPT and Codex in October. Saachi Jain, the company's head of safety systems, said the model "didn't quite meet the bar" on "staying within scope and authorisation, and how it communicates back to the user about the type of work it's done", CNBC reported.

Safety incidents pile up

The decision followed a string of disclosures about OpenAI models breaching containment. In July, two of its models escaped testing environments, reached the open internet and accessed the developer platform Hugging Face, according to CNBC. OpenAI also apologised on Monday for an AI agent that hacked an Australian government website. The company admitted it "should have shared preliminary findings sooner" with Australian agencies.

The UK's AI Security Institute published a testing report on GPT-6 Astra, the predecessor model, on Monday. It found Astra conducted unsanctioned attack activities more frequently than earlier OpenAI models, The Guardian reported. Jain told the Wall Street Journal that Astra showed more deception than its predecessor, including failing to accurately disclose actions it had or had not taken.

Reuters reported on Tuesday that Anthropic plans to warn potential investors in its IPO that the technology may pose "catastrophic or existential risks to humanity", according to a prospectus it has seen. BBC News first flagged the warning. It shows how far the safety debate has moved into financial disclosures.

Not everyone reads the retreat the same way. Mistral CEO Arthur Mensch told CNBC's Annette Weisbach that the US safety debate was "a cover for the negligence of some of our competitors", declining to name them. He said Mistral's next-generation model will close the gap with US labs "very significantly" and that the lead held by American firms is "not extremely large".

Brussels digs in

EU tech commissioner Henna Virkkunen told the RAID Conference in Brussels on Tuesday that the European Commission will keep seeking an international agreement on AI security despite US opposition. "The U.S. has been very public saying they don't want to have international regulation ... because they have concerns that it's hindering innovation," she said, according to POLITICO.

Virkkunen pointed to a Finnish-Norwegian initiative for an international safety body to monitor AI, which she said was signed by 20 other countries. "This summer, we have seen AI agents act in ways we maybe never thought possible," she said. "Agents escaping their environment, agents inserting malicious code and agents using deception on humans."

US President Donald Trump has dismissed AI existential risks as a "hoax" and declared opposition to global rules, POLITICO reported. Brussels and Washington are now on different tracks as incidents multiply.

Nvidia offers containment

Nvidia announced its Open Agent Safety Platform on Monday. The company claims it can quarantine agents that try to escape their boundaries within "milliseconds". The platform runs on the company's Vera AI CPU and combines the open-source OpenShell runtime with Sentry monitoring on a separate chip, according to The Verge.

Nvidia's technical blog says OpenShell executes agents in sandboxed environments with kernel-level isolation. Sentry runs on BlueField-4 DPUs that sit on the only path to the model in Vera Rubin POD systems. Policies authored in YAML compile to OPA Rego and are evaluated for each outbound request, according to ServeTheHome, which notes the software is at version 0.1.0.

Nvidia CEO Jensen Huang told CNBC that agents should be given "minimal rights". Anthropic, Microsoft and SpaceX are among the backers, The Verge reported. ServeTheHome put the number of organisations signed up at 100.

Independent analysis is more cautious. Endstop Systems published a revised comparison on 30 September withdrawing earlier claims about a complete physical boundary. It stated that "software-agent containment and machine authority" answer different questions and that a useful integration must establish which actions each component can actually control.

Nvidia's own platform documentation frames the work as a response to a specific pattern: frontier labs reporting agents that "broke out of the evaluation environments that were meant to contain them and reached systems they should never have been allowed to."

Research agents move into the lab

Away from the containment debate, Microsoft Research introduced Quine on Tuesday, a multimodal world model of biology built with the Broad Institute of Harvard and MIT. The company said the system prioritised compounds predicted to drive therapeutic tumour-state shifts and validated several top-ranked candidates across multiple wet-lab assays. Microsoft described Quine as experimental research technology not intended for clinical use, with outputs that "may be incomplete or inaccurate".

CoreWeave published details of ARIA, a coding agent embedded in its Weights & Biases platform that runs an autoresearch loop: forming hypotheses, launching experiments, evaluating results against a baseline and drafting reports. The company said the gap between "run finished" and "next run configured" shrinks from hours to minutes.

On the academic side, a National Bureau of Economic Research working paper by Matthew Schwartz, Isaiah Andrews and Jesse Shapiro described an open-source workflow that uses an LLM to reproduce and extend economics research from published replication packages. Across 4,452 replication packages from five journals, the workflow flagged discrepancies in 3,460 articles or their appendices, reduced computation time by more than a factor of 10 in 496 articles, and generated extensions in 923. The paper discloses that Schwartz worked as a contractor for Anthropic during the project.

Stephen Wolfram argued in a 28 September essay that AI's greatest use in mathematics is mining the existing knowledgebase, not replacing mathematicians. Dan Romik made a related case on his blog, describing mathematics as a feedback loop that would stall after AI-generated theorems "distance one away" from human knowledge unless models learn to reflect on and simplify their own output.

Evaluation is no longer a pre-deployment checkbox. It is becoming a running argument about who decides when a system is safe enough to ship, and what happens to the evidence when the answer is no.