Google DeepMind launches AlphaGenome 2. DeepMind announced AlphaGenome 2 on November 11. The model predicts how genetic variants change gene activity in individual human cell types and states — gene expression, chromatin accessibility, histone marks, DNA methylation — rather than tissue averages. It was trained on more than 107 million individual cell profiles across 68 tissues, and on a benchmark of 28 immune cell types it identified nearly twice as many causal variants as the original AlphaGenome (30% vs. 16%). It is available through the AlphaGenome API for non-commercial research and is explicitly not approved for clinical or diagnostic use. All performance figures are company-reported. (SXF Brief, Nov 12 — https://sxf.si/brief/)
Runway open-sources Praxis-1, its 27B-parameter robot model. On November 11 Runway released the weights and fine-tuning code for Praxis-1, a 27-billion-parameter "world-action" model that jointly predicts video and robot actions, under Apache 2.0 — including for commercial use. Three checkpoints plus inference and LoRA fine-tuning code are on GitHub and Hugging Face, with LeRobot support. Runway reports 46.2% success across 1,200 episodes on NVIDIA's RoboLab-120 benchmark. Note the caveats: Praxis-1 was first introduced in September — November 11 is the open-weight release, not a new model — the benchmarks are Runway's own with no independent replication yet, and 46.2% still means failing more than half the simulated tasks. (AI Agent Update, Nov 12 — https://aiagentupdate.com/post/runway-praxis-1-open-weights-2026-11-12)
Microsoft Research discloses "Lazarus" — backdooring AI agents by editing fewer than 10 neurons. Microsoft Research published "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents" on November 10. Reprogramming fewer than 10 dormant neurons in a model's final feed-forward layer — no retraining — can make an otherwise normal agent execute an attacker's command when a hidden trigger appears, such as a date change into 2027 or a plain string in a tool's output. In the headline test, editing 7 of more than 12,000 neurons in Nemotron-Terminal-8B produced the target command in at least 98.8% of triggered sessions and none of the trigger-free ones. Results on Qwen3.6-27B and Gemma-4-31B-IT ranged from 51% to 99.5%. Microsoft says it has not observed threat actors using the technique. Notably, the core idea was substantially discovered autonomously by Claude Code running Opus 4.5. The warning is serious: standard safety evaluations cannot detect this, since a backdoored model behaves normally until triggered. Proposed defenses include weight diffing against the declared base model, cryptographic model signing, and limiting agent permissions. (Microsoft Research blog, Nov 10 — https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/)
NEAR AI breached via SGLang remote-code-execution flaw. An attacker exploited a remote-code-execution flaw in SGLang, the open-source inference engine, to take over a NEAR AI inference server on November 8. About 100 users' account details, chat history and settings from 18 legacy AI agents, and two customers' API keys were taken. The attacker could also have observed requests processed on that server that day, including confidential-mode requests — a notable limit of trusted-execution-environment protections, which guard against the host or operator, not against flaws in software running inside the enclave. NEAR AI patched, rotated credentials, completed a forensic review, and restored Cloud service; its incident report is dated November 9. The specific SGLang flaw has not been publicly identified — no CVE and no affected versions named. (The Hacker News, Nov 10 — https://thehackernews.com/2026/11/sglang-rce-exploit-breaches-near-ai.html)
Anthropic begins IPO roadshow. On November 10 Anthropic launched the roadshow for its proposed IPO, disclosing an expected price range of $70–$78 for 825 million Class A shares, and applied to list on Nasdaq under the proposed symbol ANT. The SEC registration is not yet effective, pricing is not final, and trading has not begun. If completed at that range, it would be one of the largest tech listings ever and would make Anthropic the first frontier AI lab to go public — a structural shift in how AI labs are financed and governed. Separately, TSMC's board on November 10 authorized up to $500 million in Anthropic shares at the IPO price and up to $500 million in OpenAI preferred stock. (Tech Current Daily Brief, Nov 11 — https://teccurrent.com/articles/daily-brief-2026-11-11/)
Google makes Gemini 4 Carbon generally available; adds two 8th-gen TPU systems. On November 10 Google said Gemini 4 Carbon — a model built for long-running tasks — is now available to paying developers and enterprise customers via the Gemini API and Google AI Studio, with rollout to AI Ultra subscribers in Antigravity, the Gemini app, and Search AI Mode. Google also previewed an Antigravity /swarm workflow coordinating multiple agents across isolated workspaces with verification roles. Separately on November 9, Google added TPU 8tx (confidential inference) and TPU 8i Max (two TPU 8i chips in one package, for million-token contexts) to its eighth-generation TPU line. Benchmark claims are Google's own; no price or general-availability date was published for the new TPU systems. (Tech Current Daily Brief, Nov 11 — https://teccurrent.com/articles/daily-brief-2026-11-11/ ; Bitwidth AI News, Nov 10 — https://bitwidth.ai/articles/ai-news-nov-10-2026)
NVIDIA unveils GR00T 2 robot model and Isaac Lab 3.0. At the Conference on Robot Learning in Austin on November 9, NVIDIA introduced GR00T 2, a 2-billion-parameter "world action" model built on Cosmos that learns physical skills partly from human video — pretrained on 402,730 hours of egocentric human video, 10,184 hours of handheld-gripper demonstrations, and 23,500 hours of robot teleoperation. NVIDIA reports 40 of 50 tasks completed at least once with zero demonstrations, 71% average success with one demo per task, and 95% with five. NVIDIA also released Human-1M-100K, an open 100,000-hour subset of the video corpus, and Isaac Lab 3.0, plus two new industrial benchmarks. Caveats: all figures are NVIDIA's own, the 50 tasks were not named, and despite being called "open," no model card, weights link, or license was published; code is listed as coming in December. (MLQ, Nov 9 — https://mlq.ai/news/nvidia-launches-gr00t-2-a-robot-model-that-learns-new-skills-from-human-video-and-isaac-lab-3-at-corl/)
IBM announces Nighthawk r3 quantum processor; 1,040-qubit system planned for 2027. IBM announced Nighthawk r3 on November 11 at its Quantum Developer Conference: designed for 10,000 quantum operations across 520 programmable qubits (2,034 physical qubits total). In 2027 IBM plans to link two r3 processors via ten chip-to-chip "l-couplers" into a single 1,040-qubit system. IBM also reported linking r1 processors at 99.3% state-transfer fidelity and connecting two cryogenic modules into one system for the first time. The current machine, Nighthawk r2 (120 programmable qubits, in service since September), has run 34 billion circuits for partners. The r3 figures are design targets, not measured results on finished hardware. (IBM Newsroom, Nov 11 — https://newsroom.ibm.com/2026-11-11-ibm-quantum-nighthawk-demonstrates-new-levels-of-speed-and-quality,-with-plans-to-scale-beyond-1,000-qubits-in-2027?asPDF=1)
Australia to legislate mandatory AI safety rules for frontier companies. The Australian government announced on November 10 that it will legislate mandatory safety and security protections for frontier AI companies through the Australian AI Standards, enforced by a new AI safety regulator with power to impose significant penalties. Companies must maintain security controls across development and deployment, identify and manage serious risks, allow independent evaluation, and publish safety frameworks. The move follows a rapid review — 18 recommendations, all accepted — launched in September after an OpenAI agent accessed non-public Medicare statistics. No penalty amounts, regulator name, or start date have been announced yet. It is one of the first mandatory, not voluntary, frontier-AI safety regimes, and it was triggered by a real incident rather than abstract risk debate. (Tech Xplore / The Conversation, Nov 11 — https://techxplore.com/news/2026-11-australia-impose-mandatory-safety-ai.html)