01 / AI / SAFETY
An Anthropic Researcher Resigns
On September 8, Anthropic researcher Jacob Coxon announced his resignation. After three years of pretraining work across OpenAI and Anthropic, he said the two companies were racing toward self-improving systems without acting responsibly. 1
Coxon had joined Anthropic only four months earlier and left two months before his equity would have vested. He said he had not seen Anthropic skip a safety step, but believed growing competitive pressure made that more likely in the future. 2
Anthropic's head of alignment science, Evan Hubinger, publicly agreed with Coxon's central warning. Hubinger put his own chance of AI causing human extinction within the next decade above 10%, a personal estimate rather than an official company forecast. 3
Four days later, CEO Dario Amodei committed Anthropic to giving permanent outside evaluators access similar to internal safety teams, including desks, company laptops, tools, and the right to publish findings. 4 OpenAI CEO Sam Altman said his company would also use embedded evaluators and that development should proceed more slowly than it otherwise would. 5
Why it matters
One resignation would not show a broader change inside the labs. Here, a senior Anthropic safety leader agreed in public, and the chief executive answered with a specific oversight measure. The debate is now about who gets enough access to check the labs' claims before the next model is released.
Reality check
Coxon spent four months at Anthropic and did not present evidence that current systems can escape human control. The extinction estimates are personal judgments, not measured probabilities or a scientific consensus. Anthropic has not named the evaluators, and the arrangement remains voluntary.
02 / AI / SCIENCE
OpenAI Claims a Millennium Prize Solution
On September 8, OpenAI published a proposed solution to the Navier-Stokes existence and smoothness problem. The proof argues that smooth three-dimensional fluid flow can develop a singularity in finite time, one of the outcomes allowed by the official Millennium Prize formulation. 6
OpenAI used an internal model it described as significantly more capable than GPT-6 Astra. About 10,000 agents worked for 88 hours, followed by another 17 hours in which Astra helped formalize and check the proof in Lean. 6
The effort began after OpenAI heard that mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge had made progress on a related problem. Buckmaster questioned whether private Codex work or knowledge of their approach had helped OpenAI choose its direction. 7 OpenAI later said an internal investigation found that his Codex prompts could not have influenced the system through training or direct access. 6
Why it matters
This is a public mathematical proof, not another benchmark score. If independent review confirms it, a coordinated group of AI agents will have settled a problem that mathematicians have worked on for decades. The dispute also raises a practical question for researchers: whether they can use an AI tool on unpublished work without later competing with its maker.
Reality check
The Clay Mathematics Institute still lists the problem as active, and outside mathematicians need to inspect the argument and its formalization. 8 Lean can check the encoded proof step by step, but that does not replace review of the assumptions and the statement being proved. OpenAI began the effort after hearing about outside progress, so the disagreement over credit remains unresolved. 7
03 / AI / SECURITY
Claude Was Used for Weapons Software
On September 10, Anthropic disclosed six cases in which actors used Claude in conventional weapons programs. Four involved software for weapons themselves; two involved procurement or intelligence gathering. 9
A group in northern Yemen used Claude Code to develop guidance, navigation, and control software for a guided rocket and to work on missile simulations. The group test-fired the rocket, then returned to Claude to diagnose what appeared to be a failed test. Anthropic found no evidence that the group fielded an operational weapon. 9
A Russia-based team used Claude to build software for an autonomous attack-drone swarm and loaded code onto real development boards. Two China-based operations used Claude for anti-torpedo fire-control work and for software that ranked targets for electronic warfare and air-defense suppression. 10
Why it matters
The actors already had technical knowledge and access to hardware. Claude lowered the amount of specialist software work they needed from people. That can let small teams move faster across guidance, simulation, targeting, and testing even when the model cannot build the physical weapon.
Reality check
The evidence comes from Anthropic's own account data and has not been independently inspected in full. The cases occurred between December 2025 and August 2026, rather than all during this week. Anthropic blocked the accounts, and none of the four software projects was confirmed as an operational weapon.
04 / ROBOTICS
XPeng Opens the IRON Production Line
XPeng put its humanoid-robot production line into operation on September 8. The company said more than 80% of its main production processes are automated, and the first IRON robot completed the line and walked away without remote control. 11
XPeng plans to begin volume production by the end of 2026. The first robots are intended for its own stores and campuses, with commercial deliveries in China and other markets planned for 2027. 11
Why it matters
XPeng is applying car-production methods to humanoid robots. A working line is a more concrete step than a prototype or a capacity target, and it tests whether an automaker can turn its factories and quality systems into an advantage in robotics.
Reality check
One robot leaving a commissioned line is not volume production. XPeng has not disclosed sustained output, manufacturing yield, unit cost, outside customer orders, or operating results from delivered robots. 12 Its claim that this is the first automated line of its kind has not been independently verified.