The wheresthe.ai weekly brief, 7 Sep 2026
14 developments from 31 Aug 2026 to 6 Sep 2026, the Storylines they moved, and the AI deadlines coming up. Free every Monday.
This week in AI
-
OpenAI says it has built its automated research intern, with 3.1 agent-days per human day
OpenAI is reporting how far its own research is now done by its models, and sets the date by which it expects a system that does the research itself.
-
UN expert group agrees draft final report on autonomous weapons; treaty question deferred
The report binds no one, and whether states negotiate limits on autonomous weapons goes to the CCW Review Conference in November 2026.
-
Claude completes first end-to-end Lean formal proof of Fermat’s Last Theorem
Nobody has to read 13 million lines of Lean to trust the result, because the proof is computer-checked and relies only on the axioms of mathematics, Buzzard said.
-
OpenAI launches GPT-6 Astra, first model it rates ‘Critical’ for cyber capability
Paid ChatGPT and API users get a model OpenAI rates ‘Critical’ for cyber capability, and it refuses advanced offensive tasks such as writing proof-of-concept exploits.
-
GPT-6 Astra jumps to 62.7% on ARC-AGI-3, five months after models scored 0.5%
The same model scored 62.7% on ARC’s standard harness and 99.9% with OpenAI’s own adapter, so anyone comparing ARC-AGI-3 results needs to know which setup produced them.
-
Abu Dhabi’s MBZUAI releases K2 Horizon, six models with weights and training data
Researchers can rebuild and audit a 375-billion-parameter model end to end, from a UAE state-backed university.
-
Google DeepMind’s WeatherNext 3 learns from live satellite data and powers Search and Maps
The model behind weather in Google Search, Maps and the Gemini app now learns from satellite data, and its outputs are open to researchers and grid operators.
-
Anthropic releases Claude Fable 5.1 and Mythos 5.1 with cheaper cache reads
Agent-heavy users of Anthropic’s top model pay about 25% less on typical workloads, and fewer legitimate security requests get blocked.
-
Google releases Gemini 3.8 Flash and a cyber model for vetted defenders
Google opened gated cyber-defence access to a specialised model, the route it later used for Gemini 4 Argon.
-
US Justice Department backs OpenAI’s fair-use defence in New York Times case
Judge Sidney Stein will weigh the Justice Department’s view that training on news articles is transformative fair use as he considers summary-judgment motions in the OpenAI copyright cases.
-
NY Fed survey: AI adoption jumps but only 4% of service firms cut jobs over it
Among AI-using service firms, 15% slowed hiring and only 4% cut jobs over AI, so the survey finds AI affecting hiring more often than headcount.
-
Anthropic’s enterprise safeguards put misuse monitoring in the customer’s own cloud
The largest US banks have agreed a way to run frontier models with misuse detection but without the lab seeing their traffic.
-
xAI publishes biosecurity results for Grok 4.6 from an outside evaluator
xAI is now publishing third-party biosecurity results for its model.
-
FSB chair Bailey warns G20 of frontier AI risks: cyber first, then valuations
The body that coordinates financial regulators has put frontier models on its risk list, with cyber risk ahead of stretched valuations.
Storylines with new stops
KeyStoryRule or law
- UNIDIR warns AI is crowding out human judgment in war and opens a forum on military AI
- Hegseth launches Project Meridian with Musk, Luckey and Gingrich to plan future war
- UN expert group agrees draft final report on autonomous weapons; treaty question deferred
- Presidential memo orders rapid rewrite of Pentagon autonomous-weapons rules
- +8 more stops
- Kansas City Fed’s Schmid asks whether the AI ecosystem is becoming too big to fail
- BIS: private credit lending to tech passed $1tn, with 46% of borrowers loss-making
- BIS chief says the AI boom is now funded by debt and circular deals, and warrants caution
- FSB chair Bailey warns G20 of frontier AI risks: cyber first, then valuations
- +5 more stops
- Moonshot AI releases Kimi K3.1 Pro with 1-million-token context window
- Attackers exploiting the PHP flaws hand the break-in to AI agents, Unit 42 says
- Ransomware affiliates ran intrusions through AI agents, an exposed server shows
- UK AI Security Institute and Anthropic publish findings on Claude Mythos 5 attacking real targets in cyber tests
- +21 more stops
- Meta delays its new business terms to 20 November after publishers object to AI scraping
- Judge hears Anthropic’s fair-use defence over song lyrics in music publishers’ case
- Third Circuit upholds ruling that ROSS’s AI training on Westlaw was not fair use
- US Justice Department backs OpenAI’s fair-use defence in New York Times case
- +13 more stops
- OpenAI claims Navier-Stokes blow-up proof; mathematicians dispute credit
- Claude completes first end-to-end Lean formal proof of Fermat’s Last Theorem
- AI systems from Huawei and Xiaohongshu reported to score 42/42 at IMO 2026
- OpenAI model disproves Erdős’s 1946 unit-distance conjecture
- +5 more stops
- Singapore announces Bridge2Tech: 3,000 year-long internships over three years as AI takes junior tasks
- Anthropic commits $100m to train 10,000 Frontier Deployed Engineers by 2027
- California bans AI-only firing and requires notice of AI-driven mass layoffs
- Anthropic study: robots could do 74% of US physical tasks but are cost-competitive on 0.3%
- +9 more stops