Today’s edition is dominated by mathematics: an Anthropic researcher says he and Claude have proved the Riemann Hypothesis, and released the proof only as machine-checkable Lean code. On the security side, Amazon documented attacker-run AI agents that found and exploited a zero-day on their own, and Signal started blocking operating-system AI agents from reading its screen. There is also a strong open-weights day, led by a 20 GB looped model from Thinking Machines.
Top 3
- Riemann Hypothesis claim: Levent Alpöge says a Claude-assisted proof, published as a Lean project, settles RH and a form of GRH; early compile checks pass, human understanding is still missing.
- AI agents found a zero-day unaided: Amazon says a China-nexus operator’s Cairn agents went from API mapping to file write in MONAI Label in about 11 minutes.
- Inkling 2 Mini: an Apache 2.0, roughly 20 GB looped Mixture-of-Experts model aimed at local agentic coding.
Research: the Riemann week
On 20 November, Anthropic number theorist Levent Alpöge posted a four-page note saying that, working with Claude, he found a proof of the Riemann Hypothesis and of the Generalized Riemann Hypothesis. He released only a Lean formalization, not a written paper, and says he has not yet mastered all of the mathematics involved. MIT’s Andrew Sutherland told Scientific American he compiled the Lean project independently; an independent review of the archive reports roughly 17 million lines, with the main theorem stated against Mathlib’s own definition and the Lean version of GRH covering Dirichlet L-functions. Critics point to timing: the note came a day after Anthropic’s IPO, and number theorist Andrew Granville called the write-up garbled. Anthropic itself has not commented. The Clay Institute still requires publication and two years of scrutiny.
The same day, Stanford’s Ravi Vakil posted a separate, all-human GRH draft by postdoc Alessandro Masullo. About eight hours later he added that several people had reported problems in it, found with the help of AI models. Why it matters: a machine-checked proof that no human yet understands is new territory; treat it as a strong claim, and watch for published axiom audits from several independent builds.
Security
- Cairn agents exploit a zero-day. Amazon Threat Intelligence found an unauthenticated Cairn server (an open-source autonomous pentest framework) used against East Asian universities. Given the goal “obtain shell access”, the agents found a path traversal in MONAI Label, a medical-imaging AI app, then wrote a reverse shell into
sitecustomize.pyand got root. Elsewhere they spent 16 hours and 148 actions on an open Redis server without success. VentureBeat reports the fix is merged but no patched release exists yet (PyPI is still 0.8.5). Action: if you run MONAI Label, enable authentication, keep it off untrusted networks, run it unprivileged, and check for altered files. - Signal blocks OS AI agents. Signal now blocks screenshots on iOS 27 and turns on screen security on Android when it detects an AI assistant reading the screen. It cannot do this on macOS and recommends adding Signal to Siri’s excluded apps there. Signal asks Apple, Google and Microsoft for a documented opt-out for apps. Why it matters: any app handling sensitive data faces the same gap; check what your mobile fleet’s assistants can read.
Models and platforms
- Inkling 2 Mini (Thinking Machines). A looped transformer: the middle 12 of 24 layers run three times per token. It has 26.4B total and 3.4B active parameters, a 1M-token context, ships natively in NVFP4/MXFP8 at about 20 GB, and is Apache 2.0. The lab reports 48.2% on DeepSWE v1.1, best among open models under 100B, and warns about repetitive reasoning loops. Ollama supports it on Apple silicon only for now. Action: worth a local test for coding agents; set output-token and turn limits.
- Tencent Hy4 open weights. The 770B-total, 49B-active multimodal MoE flagship with a 1M context is now downloadable, with vLLM and SGLang recipes and a system card. TechNode reports Apache 2.0, while the GitHub repo lists its licence as “Other”. Action: read the licence file before commercial use.
Agents, automation and tools
- Claude Code 2.1.330. Adds sign-in to organizations managed in Anthropic’s Enterprise Admin Console, and a built-in chart mod. It fixes freezes after compacting long conversations and on some WebFetch pages, a worktree exit that could delete another session’s worktree, and lost settings when several processes save at once. Action: update if you run parallel sessions or worktrees.
- AWS STS GetScopedCredentials. A new API that narrows existing temporary credentials for an agent or sub-agent. The result is the intersection of the session’s permissions and an inline policy, keeps the caller’s identity, is logged in CloudTrail, and can be narrowed again (up to three policies). Action: stop passing full role credentials to agents; scope them per task.
AI tools of the day
- DreamCoder: a browser-based coding harness and IDE with a tiling window manager, from Mickey Petersen of Mastering Emacs, announced after 15 months of work. Announcement
- AWS DevOps Agent Tools: AWS-curated skills, custom agents, MCP servers and partner connectors for DevOps Agent incident work; Apache 2.0, about 143 stars. GitHub
- Spring AI 2.1 GA: supports the stateless MCP 2026-07-28 revision, adds a Session API, tool-call approval and fixes for 13 CVEs (also in 2.0.3). Release post
Open source and local AI
- DaisyRose-27B (RadixArk). An agentic coding model post-trained from Qwen3.8-27B with reinforcement learning only, Apache 2.0, with FP8 at about 29 GB and GGUF builds. Self-reported Terminal Bench 4.1 rises from 5.6 to 12.0. Action: wait for independent benchmarks.
- NVIDIA Nemotron-Labs-Sigma. 14 open checkpoints (3B and 8B) of a continuous diffusion language model that denoises text in an embedding space, under the Nemotron Open Model License; GPU only. Why it matters: a research-grade look at non-autoregressive text generation, not a production model.
Worth watching
- Scientific American: Riemann Hypothesis “solved” by Claude: a short explainer of why the claim matters, if it holds.
- Abundant Future: The Riemann Hypothesis proof, explained: a plain-language walk through how Lean checking works and what it does and does not certify (the video’s claims about other checks are not confirmed here).
Sources
- Nature: Mathematician says he and AI solved Riemann hypothesis
- Scientific American: Anthropic mathematician claims to prove RH
- The Verge: Riemann claim
- Tao Nhu: Inside the Lean archive
- Ravi Vakil: Hurried Announcement, Masullo GRH
- AWS Security Blog: Cairn AI agents found a zero-day
- VentureBeat: no patched MONAI Label release
- Signal: Signal is Serious About Protecting Your Privacy
- Thinking Machines: Introducing Inkling 2 Mini
- Inkling 2 Mini model card
- Ollama releases
- TechNode: Tencent releases Hy4 weights
- GitHub: Tencent-Hunyuan/Hy4
- Claude Code releases
- AWS Security Blog: GetScopedCredentials
- Mastering Emacs: Announcing DreamCoder
- AWS: DevOps Agent tools repository
- GitHub: aws/tools-for-devops-agent
- Spring AI 2.1.0 GA
- MarkTechPost: DaisyRose-27B
- MarkTechPost: Nemotron-Labs-Sigma
- YouTube: Scientific American
- YouTube: Abundant Future
