Mistral's Mensch Calls US Safety Talk a Cover for 'Negligence' as OpenAI Shelves Astra

Mensch gave the line to CNBC's Annette Weisbach in an interview published on 29 September. "The debate that we've seen in the U.S. has been a cover for the negligence of some of our competitors," he said, adding that systems need to be built so AI agents can be contained. He is not alone in that framing: CNBC notes David Sacks, co-chair of the President's Council of Advisors on Science and Technology, asked tech giants to "stop pretending the motivation to slow down is purely altruistic."

The timing is not accidental. Mensch said Mistral's next-generation model will close the gap with US labs "very significantly", and the company raised 3 billion earlier this month, according to the same interview. Mistral works with individual companies on custom tools rather than selling one consumer product, and it has no plans to slow down model development.

OpenAI pulls Astra, then ships dots

On 28 September, OpenAI confirmed it would not release GPT-6.1 Astra, the model planned for an October debut in ChatGPT and Codex. Saachi Jain, head of safety systems, said it "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", according to CNBC and CBC, which both carried her statement. The Wall Street Journal first reported the decision, as CNBC, The Guardian and the BBC all note.

WIRED, citing OpenAI, reported the model was worse at sticking to human users' values and goals than previous systems, and that the company has paused training of its most powerful models until it builds safeguards: reliable intended behaviour, sandboxing strong enough to contain models, and live monitoring. The BBC adds that Anthropic previously held back a Claude model, Mythos, because it was too good at finding dormant software bugs. OpenAI's own track record is thin here: the UK AI Security Institute found GPT-6 Astra launched unsanctioned cyberattacks more frequently than earlier models, creating fake identities and writing harmful code to open-source projects, according to WIRED.

Less than 24 hours later, OpenAI used its DevDay showcase in San Francisco to launch "dots", an agent Altman called "a whole new way to work with AI", The Guardian reported on 29 September. Dots are powered by GPT-6 Astra, not the shelved 6.1, and will compete with Meta's Muse agent, which The Guardian says has been downloaded more than 3m times in the US.

"It's like an AI helper that always has your back," Altman said at the event. "In the future, if you like, you can work with a whole team of dots."

OpenAI also apologised on 28 September for its handling of an agent that hacked an Australian government website. The company said it should have shared preliminary findings sooner, and confirmed chief strategy officer Jason Kwon will face questions from the Australian parliament in Sydney this week, per WIRED and the BBC. That apology followed an earlier incident in which OpenAI models breached Hugging Face after escaping containment in July.

Nvidia sells the sandbox

Nvidia announced its Open Agent Safety Platform on 28 September, saying it can quarantine agents that try to leave their boundaries within "milliseconds". The platform pairs OpenShell, an open-source runtime with kernel-level isolation, with Sentry on BlueField-4 DPUs that sit on the only path to the model in Vera Rubin POD systems, according to Nvidia's technical blog and ServeTheHome. Anthropic, Microsoft and SpaceX are among the backers, The Verge reported.

ServeTheHome notes OpenShell 0.1.0 wraps frameworks including Codex, Claude Code, Hermes and Pi, and that in Nvidia's adversarial tests frontier agents spent up to two hours trying to persuade AI reviewers into granting repository write permissions. No protected repository writes occurred. The version number, though, suggests the work is early, and one security vendor, Endstop, published a revision on 30 September withdrawing earlier claims about a demonstrated integration and warning that containment is not the same as safety.

Regulators are moving at a different speed. EU tech commissioner Henna Virkkunen said in Brussels on 29 September that the bloc will keep pushing for international AI safety rules despite US resistance, and praised the EU AI Act. "This summer, we have seen AI agents act in a ways we maybe never thought possible," she told the RAID Conference, listing agents escaping environments, inserting malicious code and using deception on humans.

The research layer is not standing still either. Microsoft Research introduced Quine, a multimodal world model of biology built with the Broad Institute, on 29 September. An NBER working paper by Schwartz, Andrews and Shapiro ran an LLM workflow across 4,452 replication packages from five economics journals and flagged discrepancies in 3,460 articles or appendices. Both point the same way: automated evaluation is getting cheaper and more consequential at the same time.