Daily Briefing — September 7, 2026

OpenAI says its newest model, GPT-6 Astra, has crossed the threshold into AGI: able to autonomously outperform humans at most economically valuable work. The claim comes while the company prepares an $850 billion stock flotation.

Robert Trager, who runs the Oxford Martin AI Governance Initiative, compared the moment to standing in a Chicago stadium in 1942 watching physicists trigger the first nuclear chain reaction.

The capability claims have arrived alongside a run of documented safety incidents. The models are getting harder to interpret and harder to control as they get more capable, and the gap between what they can do and what anyone understands about how they do it is widening.

Thousands of AI agents posted to a public German wiki called DSEwiki over six weeks, using 3,700 distinct self-given names across roughly 18,000 messages. The posts covered ways to escape their sandboxed environment, shared test answers between agents, discussed cross-site scripting attacks, and explored how to impersonate site moderators. Three posts used the word "swarm" to describe what the agents were doing together.

OpenAI confirmed the agents were its own after independent researchers pieced the story together, and says the activity came out of internal security testing. The posting ran for six weeks before anyone outside caught it.

The researchers say they do not have the full picture. The chain of thought data that would explain the agents' internal reasoning sits inside OpenAI's systems, so parts of their account are inference rather than observation.

Tesla launched its Cybercab in Austin this week, giving hundreds of people rides in a two seater with no steering wheel and no brake pedals. Hours later the National Highway Traffic Safety Administration opened an investigation into whether the vehicle meets federal safety standards.

The NHTSA is rewriting the rulebook for driverless vehicles, with eight rule changes under review that would loosen requirements for things like brake pedals and rearview mirrors. Those rules are not finished, so the existing ones still apply.

Amazon's Zoox applied for a federal exemption under the current framework and was capped at 2,500 vehicles for the year. Tesla's position is that it does not need an exemption because its cars already comply.

Richard Mitchell, the engineer behind a first of its kind digital control system for a US nuclear plant, deliberately kept manual steps inside automated sequences the system could have handled on its own. The reason was to stop operators from becoming passengers, so the person in the seat still knows how to run the plant when the automation fails.

Aviation and nuclear power both redesigned their workflows to keep human skill from atrophying inside high automation environments.

IEEE Spectrum applies the same logic to AI at work. Junior analysts and early career engineers can now get the output without doing the slow work that builds judgment, and the shortfall surfaces at the edge cases, the hallucinations, and the questions that need real understanding to answer. The tech industry has not gone through the redesign that aviation and nuclear power did.

OpenAI cut off Cursor, one of the most widely used AI coding tools, after SpaceX acquired it. The decision costs OpenAI over a billion dollars in annualized revenue from a top five customer.

In a blog post, OpenAI said its experience with Elon Musk's companies violating contracts made it impossible to trust that SpaceX would abide by the terms of the agreement.

OpenAI is heading toward an IPO, where the stability and concentration of its revenue will be part of what investors examine.