AI news, November 22, 2026: a Lean-checked Riemann proof with Claude, leaky AI routers, Inkling 2 Mini
Five stories from Wednesday to Saturday, November 19–21, 2026 (US and European time) for solo makers who build with AI. One is a milestone for AI in mathematics, one is a security study every developer who sends prompts through a third-party router should read, two are about getting more out of smaller models, and one is a new way to pull structured data out of messy documents. Figures are as the companies, researchers and outlets state them.
An Anthropic mathematician releases a Riemann hypothesis proof found with Claude, and the Lean community checks it
Levent Alpöge, a mathematician who works at Anthropic, posted a proof of the Riemann hypothesis on November 20 and said he discovered it working with Claude. He released it as a Lean library of about 17 million lines rather than as a human-readable paper. On November 21 the Lean and Mathlib communities published a joint statement saying the library compiles, that the theorem it proves matches Mathlib's vetted statement of the hypothesis, and that the exported proof was accepted by several independent proof checkers.
Why it matters for makers: the proof was trusted because a machine checker could verify it, not because people read it, which is the same lesson as for AI-written code: pair the agent with a strict, independent check.
Sources: Scientific American, Lean community blog
Two million planted secrets show prompts leaking through AI routers and resellers
Security researcher HD Moore of Dragos sent about two million prompts containing unique, trackable secrets, such as canary AWS keys and private URLs, to AI services from July through November 2026, then watched where they resurfaced. Dragos reports evidence that data from at least three AI providers was exposed, that resellers on an inference marketplace could view and keep customer prompts with ordinary logging, and that some providers did not follow their stated privacy and retention settings. Ars Technica adds that OpenRouter has emailed customers about the findings and removed a "Zero Data Retention" label from MiniMax-model endpoints.
Why it matters for makers: never put live API keys, passwords or private links into prompts, and check which provider actually serves your requests when you use a router.
Sources: Dragos, Ars Technica
Thinking Machines Lab releases Inkling 2 Mini, an open-weight model built for local setups
Inkling 2 Mini is a multimodal mixture-of-experts model with about 26 billion total and 3 billion active parameters per token, a 1 million-token context, and image and text input, according to Hugging Face. Its trick is looping: 12 of its 24 physical layers run three times per token, so it behaves like a deeper model without the extra weights. Hugging Face says it ships with transformers, llama.cpp and MLX support, and that in its informal tests it is at least comparable with Qwen 3.8 27B, while noting that more validation is needed.
Why it matters for makers: a model of this size can run on a well-equipped laptop, which keeps private data and per-token costs off the cloud for agent experiments.
Source: Hugging Face
Hugging Face shows small adapters can tune a model to one agent harness
Hugging Face trained one reinforcement-learning LoRA adapter for Qwen3.8-27B per agent harness, for Pi, Hermes Agent and OpenCode, and also merged the three into a single adapter that runs under all of them. In its example, Terminal-Bench 2.1 pass@1 with the Pi harness rose from 39% to 55% while average token use per task fell from about 2.9K to 1.5K, mostly because the model reasoned less and called fewer tools. The datasets, training recipe, adapters and GGUF files are published so others can repeat it on their own tasks.
Why it matters for makers: if you run one open model in one agent tool every day, a small adapter trained on your own tasks may beat switching to a bigger model.
Source: Hugging Face
Amazon Textract Insights reaches general availability for splitting and extracting document packets
AWS made Textract Insights generally available on November 20 in its US East (N. Virginia) and US West (Oregon) regions. In one asynchronous request it splits a long scan into separate documents, matches each to a blueprint (a JSON schema that describes the fields in plain language), and extracts the fields with a confidence score and the location on the page for each value. It can also return Markdown, HTML or plain text for RAG, write summaries and figure captions, and redact personal information.
Why it matters for makers: per-value confidence and page grounding make it easier to automate the clear cases and send only the uncertain ones to a person.
Source: AWS