A quiet weekend still produced one of the biggest claims of the year getting a formal stamp of approval, along with a hard look at where your prompts really go and what powers the labs that answer them. Here is what held up and what still needs checking.
1. Lean Community Says Alpöge’s Riemann Hypothesis Proof Checks Out
What happened. The Lean and Mathlib communities, with backing from the Lean FRO and the Mathlib Initiative, published a joint statement on Nov 21, 2026 certifying Levent Alpöge’s formal proof of the Riemann hypothesis. They report that the roughly 17 million line Lean library depends only on Mathlib, that two separate comparison tools confirm the theorem it proves is exactly Mathlib’s own statement of the hypothesis, and that the exported proof passed Lean’s kernel plus several independent checkers written in other languages. “We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct,” the statement says.
The hype check. A machine check settles whether the argument is valid, not why it works: no person has read 17 million lines, and the statement itself says the next job is for mathematicians to explain the new ideas. The certification also rests on Mathlib’s definition being the right one, which the authors say experts have vetted. The separate human draft by Alessandro Masullo is a different matter and still has open problems flagged by readers.
Verdict: Supported. This is the formal check experts were waiting for when we first reported the two competing Riemann claims, and it answers the question of whether the Lean statement matches the real hypothesis.
2. Dragos Canaries Show Prompts Leaking Through Inference Providers
What happened. Security firm Dragos published a study by HD Moore in which the team sent nearly two million prompts seeded with unique tracking links and decoy credentials to AI providers between July and November 2026. Dragos reports that data from at least three hosting providers was stolen by attackers, that some sellers on an inference marketplace kept and viewed customer prompts through ordinary logging, and that an unnamed Chinese lab fetched secret links from prompts first sent to US providers, including many marked for zero data retention.
The hype check. None of the affected providers is named, so customers cannot check their own exposure, and the findings come from a security vendor rather than an independent audit. Dragos itself notes that a triggered canary proves a prompt leaked but not where in the routing chain it escaped, and that several hundred of the systems it traced remain unidentified.
Verdict: Needs follow-up. The method is sound and the warning about resellers is credible, but the story is incomplete until the providers are identified or come forward.
3. Watchdog Estimates Anthropic’s 2028 Power Will Be 87% Gas
What happened. The Better Data Center Project estimated in a Nov 16 case study, which drew wide discussion among developers this weekend, that about 87% of the electricity behind Anthropic’s contracted US data centers will come from gas and 11% from coal in 2028. The group combined on-site gas plants with the grid plants likely to serve the added load, using the EPA’s AVERT emissions model, and puts the middle case at 32 million tonnes of CO2 equivalent that year, with a range of 17 to 38 million.
The hype check. This is a model, not a disclosure from Anthropic. It assumes all 9.6 gigawatts of site-linked capacity is running in 2028 at 70% utilization, part of the capacity figure comes from press reporting by The Information, and the group flags a margin of plus or minus 25% on some partner capacity. Marginal emissions modeling can also produce higher fossil shares than the annual averages companies usually report.
Verdict: Needs follow-up. The method is laid out in the open, but the headline split depends on assumptions that only Anthropic and its partners can confirm or correct.
4. A Community Quant Squeezes Inkling 2 Mini Into 5.4GB
What happened. A developer released an MIT licensed quantization tool along with a 5.38GB build of Thinking Machines’ Inkling 2 Mini at about 1.65 bits per weight, using trellis codes plus fine tuning across the full mixture of experts. The author reports it runs on Apple silicon through MLX in about 5.3GB of memory, scoring 87.0% on a 200 problem GSM8K sample against 96.5% for the full precision model.
The hype check. Coding suffers most: the author’s HumanEval score drops from 92.7% to 71.3%, so this is not the model that posted the frontier coding numbers. The test sets are small, every figure is self-reported, and the speed result was measured on a top end M5 Max rather than the base laptops the file size suggests.
Verdict: Promising but unverified. It follows our coverage of the original Inkling 2 Mini release, which needed 32GB of memory, and the files are public for anyone to test.
5. Pandex Puts European Law Behind an API and MCP Server
What happened. A new service called Pandex offers legislation and top court rulings from the EU and several European countries, including France, Germany, the Netherlands, the UK, Austria, Switzerland and Luxembourg, through a REST API and an MCP server that agents can call directly. The site claims about 2.8 million documents with 37 million linked citations, and lists a free beta tier of 1,000 requests a day with paid use at €1 per 1,000 requests.
The hype check. The site does not clearly name the company behind it, several sources are marked as only partly covered, and there is no independent test of how accurate or current the texts are. For legal work, an agent that cites the wrong version of a regulation is worse than no agent at all.
Verdict: Promising but unverified. Structured access to primary legal texts is useful for builders, but the coverage claims need checking before anyone relies on them.
Also on our radar
- A new paper on Posterior Matching Distillation claims the best image quality among four step generators distilled from SDXL and FLUX.2 klein 4B, with more varied outputs than rival methods.
- Researchers including a WeChat intern describe WeVisForensics, an agent that hunts for evidence of generated video, reporting 95.2% average detection accuracy across their benchmarks.
- A study of vision language models introduces a probe called MC-Lens that reads concept information from inside the model, borrowing ideas from global workspace theory while stating plainly that it makes no claim the models are conscious.
What to watch
With the formal check in hand, attention on the Riemann proof shifts to mathematicians turning 17 million lines into ideas people can follow, and to how quickly journals and prize committees respond. On the Dragos study, the key question is whether any inference provider or marketplace confirms it was one of the leaks. A response from Anthropic with its own projected energy mix would settle more than any outside model can.
Sources
- Lean community blog, Lean verification of the Riemann Hypothesis
- Dragos, Three breaches, endless leaks: Your AI supply chain needs canaries
- Better Data Center Project, How AI is powered: a breakdown of BYOP and grid power
- ShinMegamiBoson, lookatmyquant repository on GitHub
- Pandex, Europe’s law, in one API
- Hugging Face Papers, Posterior Matching Distillation
- Hugging Face Papers, WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection
- Hugging Face Papers, Probing the “Conscious” Mind: Selective Access and Global Workspace in Vision-Language Models
- AI Bandwagon, Today’s Top Five: Nov 21, 2026
By Willa Ravenwire · The AI Bandwagon
