Runway Releases Praxis-1 Open Robot Model Weights and Training Code Under Apache 2.0
Runway opens Praxis-1's 27B-parameter world-action model weights and fine-tuning software, reporting 46.2% success on RoboLab-120.
Source: Runway Official Release
CURATED AI NEWS & ANALYSIS
SXF Brief
AI News — November 12, 2026: AlphaGenome 2, Praxis-1 & Qwen4-Max
Five stories. One clear picture.
01 / AI Genomics Research
Google DeepMind Launches AlphaGenome 2 for Cell-Specific DNA Variant Prediction
AlphaGenome 2 predicts DNA regulation in specific cell types using more than 107 million training profiles; research API access is non-commercial, not clinical.
Alibaba Cloud Opens Qwen4-Max-Preview With Multimodal Reasoning and Token Plan Access
Qwen4-Max-Preview enables text, image and video understanding, reasoning and tool calls, but is Token Plan-only and scheduled to retire November 25.
Microsoft Research Finds Lazarus AI Agent Backdoors Hidden in Few Model Neurons
A Microsoft study shows weight-level changes in fewer than 10 dormant neurons can trigger agent misbehavior while ordinary evaluations appear normal.
NVIDIA Contributes OpenShell Agent Runtime to CNCF for Kubernetes Security
NVIDIA is contributing OpenShell, an open-source security runtime for autonomous agents, to the CNCF Sandbox ecosystem.
AI NEWS ANALYSIS
Today's Big Picture
AI's next frontier is moving from impressive model outputs to deployable, governable systems. Google DeepMind's AlphaGenome 2 targets cell-specific genetic mechanisms, while Runway's public Praxis-1 weights allow robot builders to adapt a video-trained action policy. Alibaba's short-lived Qwen4-Max preview shows that access conditions matter as much as model capability. Microsoft's Lazarus research exposes hidden supply-chain risk in checkpoints; NVIDIA's OpenShell contribution tackles permissions and governance for tool-using agents. Across science, robotics and cloud operations, the decisive test is whether outputs are reproducible, access is accurately represented, and systems can be constrained in practice.
What to
Watch Next
Independent reproduction of AlphaGenome 2 variant effects in specialized cell populations
Real-robot evaluation and safe fine-tuning of the newly open Praxis-1 checkpoints
Qwen4-Max preview retirement on November 25 and any supported successor offering
Reproducible Lazarus auditing and CNCF OpenShell security testing beyond vendor claims
AI News Archive
Read previous source-backed editorials and dated briefings.
Have an AI story worth covering?
Send an AI model release, research finding or industry announcement for independent editorial review.
Pitch a Story →IN-DEPTH AI NEWS / NOVEMBER 12, 2026
Five AI News Reports and Technical Analysis
Independent reports on AlphaGenome 2, Runway Praxis-1 open weights, Qwen4-Max preview, Lazarus weight-level backdoors and NVIDIA OpenShell's CNCF contribution.
Google DeepMind Launches AlphaGenome 2 for Cell-Specific DNA Variant Prediction
AlphaGenome 2 predicts DNA regulation in specific cell types using more than 107 million training profiles; research API access is non-commercial, not clinical.
Google DeepMind introduced AlphaGenome 2 on November 11, shifting DNA-sequence prediction from broad tissue averages to the behavior of specific cell types and cell states. A variant that matters in an immune cell may do little in neighboring tissue, making cell resolution crucial for disease research. AlphaGenome 2 conditions predictions on a cellular profile alongside a DNA sequence, allowing scientists to probe regulatory mechanisms that can be hidden by tissue-level averages. This is a significant scientific-model release, but it is not a diagnostic tool or a demonstrated medical treatment.
AI Genomics Research: verified source facts
- DeepMind says its training combined human and mouse molecular data with more than 107 million individual cell profiles across 68 tissues.
- On the TenK10K immune-cell dataset, the company reports identifying 30% of causal expression variants versus 16% for the compared original system, and 43% versus 23% for directional gene effects.
- AlphaGenome 2 is available through the AlphaGenome API for non-commercial research; commercial research is restricted to private preview. DeepMind explicitly says the systems are not approved for clinical or diagnostic use.
Read more: SXF editorial analysis
How cell conditioning works in practice
The first AlphaGenome predicts many genomic readouts from a sequence across a fixed collection of tissues. Its successor takes an additional representation of the target biological condition, then predicts expression and regulatory activity for that state. In an analysis pipeline, researchers can select a cell type, score candidate variants, compare predicted shifts in expression, and prioritize laboratory follow-up. The model can also condition on a new single-cell RNA sequencing snapshot to study previously unseen cellular states without a conventional full retraining run. Carefully record the reference genome, cell annotation and variant normalization; apparent precision can be undermined by mismatched biological inputs.
SXF analysis: connecting association to biological mechanism
Genome-wide association studies frequently identify stretches of DNA associated with disease but cannot identify the exact cell where a variant acts. Condition-specific prediction makes the bridge between genetic association and molecular mechanism more actionable, especially in immunology and drug-target prioritization. The stronger claim is not simply better benchmark scores but a new way to interrogate cellular context. Researchers benefit most if predictions are paired with independent CRISPR perturbations, external cohorts and uncertainty estimates. The release is distinct from the robotics and formal-verification Google stories covered in the prior issue.
Runway Releases Praxis-1 Open Robot Model Weights and Training Code Under Apache 2.0
Runway opens Praxis-1's 27B-parameter world-action model weights and fine-tuning software, reporting 46.2% success on RoboLab-120.
Runway made Praxis-1 generally downloadable on November 11, releasing the weights and supporting code for its video-trained world-action robot model under Apache 2.0. The system was announced in September and tested with selected partners, so the new event is public open-weight availability, not its first invention. Praxis-1 predicts an upcoming video segment and the actions a robot should take, combining large-scale video learning with embodiment-specific control. The transition from restricted partners to reusable checkpoints matters for robotics developers whose hardware and tasks differ from the systems used in proprietary demonstrations.
Open Robotics Models: verified source facts
- Runway describes Praxis-1 as a roughly 27-billion-parameter mixture-of-experts diffusion transformer predicting both future video and robot actions; its research report describes roughly 2.8 billion active parameters per token in the video backbone.
- Runway reports a 46.2% success rate over 1,200 simulated episodes on NVIDIA RoboLab-120, compared with 42.9% for the cited FLUX 3 Action baseline.
- The company says weights, inference code and full or LoRA fine-tuning code are available under Apache 2.0 through Hugging Face and GitHub, with LeRobot support and integration paths including vLLM-Omni.
Read more: SXF editorial analysis
What it takes to run a policy on new hardware
Open model weights do not automatically produce safe robot motion. Developers must bind camera observations and joint state to the model's expected format, select an action head for the robot embodiment and often fine-tune on local demonstrations. The technical report lists around 54.4 GB for bf16 model weights and discusses 80 GB GPU deployment; separate inference and runtime costs remain. Test instruction disambiguation, collision handling, gripper timing and emergency stops with a guarded physical setup. A simulator can screen candidate policies but should never replace staged real-robot validation with strict operator oversight.
SXF analysis: robot models become an editable infrastructure layer
The licensing and reproducible deployment components may be more important than any single leaderboard position. They allow integrators to inspect, adapt and serve a shared policy rather than relying on a black-box endpoint. That could expand experimentation across industrial arms, low-cost research robots and warehouses, while distributing responsibility for physical safety to downstream teams. Runway's own report acknowledges problems with look-alike targets, vague instructions and long multi-step tasks. This story covers Runway's November 11 public release, not NVIDIA's earlier GR00T 2 launch or Google's earlier Gemini Robotics 2.5 pilot.
Alibaba Cloud Opens Qwen4-Max-Preview With Multimodal Reasoning and Token Plan Access
Qwen4-Max-Preview enables text, image and video understanding, reasoning and tool calls, but is Token Plan-only and scheduled to retire November 25.
Alibaba Cloud began a time-limited Qwen4-Max-Preview on November 11, providing developers with access to a new flagship multimodal model for reasoning, coding and tool-using workflows. The launch is notable precisely because its release conditions are narrow: this is a preview hosted in Model Studio, not open model weights or a permanent pay-as-you-go API. Teams evaluating it should treat the deadline and procurement channel as primary engineering constraints. The preview window makes early benchmark results useful for exploration but inappropriate as the sole foundation for a long-lived production dependency.
Frontier Model Preview: verified source facts
- Alibaba identifies the model ID as qwen4-max-preview and documents text generation, image and video understanding, reasoning and tool calling, with structured outputs.
- The official release list places the preview on November 11, 2026, and the documentation states availability through Token Plan only, with no pay-as-you-go access.
- Alibaba specifies a preview beginning at 10:00 November 11 and ending at 24:00 November 25, 2026, in UTC+8, with retirement after the stated interval.
Read more: SXF editorial analysis
How API teams should test a temporary model
Build a model adapter that isolates endpoint credentials, response formats and errors from application code. Evaluate complex tool-call sequences, schema-constrained output, multilingual prompts and multimodal document interpretation with a fixed test set. Track task success and total billed token-plan consumption rather than assuming a public per-token dollar rate exists. The documented compatibility with OpenAI- and Anthropic-style interfaces can lower integration effort, but input limits, tool semantics and rate policies must still be checked against Alibaba's current reference. Implement a fallback model before the November 25 retirement, and version stored results so they remain interpretable after the preview disappears.
SXF analysis: an API preview is a research opportunity, not availability parity
A preview can be significant if it exposes a genuine leap in reasoning and task orchestration, yet its restricted channel separates it from generally available production models. The main beneficiaries are evaluation engineers and product teams comparing model strategies, not organizations needing guaranteed long-term support. Alibaba's careful documentation of retirement is important evidence that should appear in headlines and buying decisions. Unlike the November 10 Meituan open-weight release already covered by SXF, this is a different provider and a hosted temporary service.
Microsoft Research Finds Lazarus AI Agent Backdoors Hidden in Few Model Neurons
A Microsoft study shows weight-level changes in fewer than 10 dormant neurons can trigger agent misbehavior while ordinary evaluations appear normal.
Microsoft Research disclosed Lazarus on November 10, a research attack demonstrating how malicious changes to a tiny number of dormant model neurons can create a concealed trigger in an AI agent. A compromised model can behave normally on standard evaluations and then change behavior when the trigger appears during a tool-using task. The finding matters because teams increasingly download third-party weights and grant agents access to terminals, repositories or internal files. This is a model supply-chain problem as much as a prompt-security problem: the harmful behavior is encoded in the checkpoint, not necessarily in a visible user instruction.
AI Supply Chain Security: verified source facts
- Microsoft's paper describes reprogramming fewer than ten neurons without conventional model retraining, using neurons in a feed-forward block as trigger detectors and token emitters.
- Its reported Nemotron-Terminal-8B experiment altered seven neurons and reached at least 98.8% trigger-trajectory effectiveness on a specified held-out evaluation; the work also reports experiments on other model families.
- Researchers report an AI-based checkpoint auditing method that flagged 91.4% of edited checkpoints with a 9.7% false-positive rate in the evaluated setup; Microsoft says it has not observed threat actors using this specific technique.
Read more: SXF editorial analysis
How to secure an agent model supply chain
Treat model artifacts as executable dependencies. Pin publisher identity, checkpoint hash and model version in an internal registry. Require approval and model-diff review for new quantizations, adapters and fine-tunes, and keep a baseline checkpoint for comparison when possible. Model-integrity checks should complement sandboxing, filesystem isolation, egress restrictions and least-privilege tool permissions. Log agent actions, require human confirmation for destructive steps and forbid arbitrary shell execution where the workflow does not need it. Standard behavioral evaluations alone cannot verify that a checkpoint has not been maliciously altered.
SXF analysis: weights deserve code-level security controls
The distinctive risk is that an agent can pass normal tests while an altered parameter structure waits for an uncommon condition. This shifts part of AI security investment toward provenance, cryptographic signing and post-download verification. The fact that an AI agent helped discover the mechanism also underscores the dual-use nature of automated research: model capabilities can accelerate both attack discovery and defense. Microsoft reports detectable anomalies, but the auditor has false positives and misses, so it cannot replace defense in depth.
NVIDIA Contributes OpenShell Agent Runtime to CNCF for Kubernetes Security
NVIDIA is contributing OpenShell, an open-source security runtime for autonomous agents, to the CNCF Sandbox ecosystem.
NVIDIA announced at KubeCon on November 10 that it is contributing OpenShell, an open-source runtime for autonomous agent security, to the Cloud Native Computing Foundation. As agents become services that execute tools and interact with infrastructure, the challenge is no longer just how a model reasons but what it is allowed to do in a production environment. An open governance path matters for teams operating Kubernetes clusters because security controls must be inspectable, adaptable and integrated across several model providers rather than tied to one vendor's interface.
Agent Runtime Infrastructure: verified source facts
- NVIDIA describes OpenShell as an open-source runtime for securing and governing autonomous agents, intended for CNCF Sandbox governance.
- The announcement connects the project to the Kubernetes ecosystem and the need to manage AI workloads consistently across large infrastructure deployments.
- The company frames the contribution alongside broader open-source infrastructure work presented during KubeCon + CloudNativeCon, rather than announcing a new foundation model.
Read more: SXF editorial analysis
What Kubernetes operators should verify
A runtime security layer should be evaluated on actual agent permissions: which files can be read, which network destinations can be reached, which credentials are exposed, which processes can spawn and how policy violations are recorded. Test it with a deliberately constrained agent against staged secret-exfiltration and unwanted-file-modification scenarios. Confirm how the project integrates with Kubernetes namespaces, service accounts, network policy, workload identity and image signing. Every guardrail needs clear failure behavior and auditability; a sandbox that silently falls back to permissive access is worse than an explicit rejection.
SXF analysis: governed agent execution becomes shared infrastructure
The contribution moves part of agent safety from model prompting into deployable platform controls and community governance. A CNCF Sandbox project may develop interoperability and broader scrutiny over time, helping security engineers avoid bespoke policies for every agent framework. But donation and project status are not the same as ecosystem adoption or hardened production maturity. The most important downstream evidence will be reproducible isolation tests, independent vulnerability reports and adoption by infrastructure operators who do not depend on NVIDIA products.