Daily AI digest
Meta’s Muse Spark Open brings a million-token context window and multimodal input to agentic coding. More urgently, attackers are targeting a critical Langflow flaw; Anthropic’s reported $16.8bn quarterly revenue also puts its planned IPO in focus.
Meta released Muse Spark Open, an open-weight model for coding and agentic work. It has 816B total and 42B active parameters, a 1M-token context window, and native text, image, video and audio support.
Why it matters Its long context and multimodal capabilities are relevant to model deployment and agent development.
Threat actors are exploiting a critical vulnerability in Langflow, an open-source platform for building AI applications and agentic workflows. Previdian reported 1,423 honeypot exploit attempts; the flaw has a CVSS score of 9.9.
Why it matters Companies using Langflow should assess whether their deployments are exposed to this actively exploited flaw.
SPACE is an open-source sandbox platform used to run tasks in Perplexity Computer. Each agent task gets a Firecracker microVM, while credential and outbound-traffic controls operate outside the VM; SPACE is available under Apache 2.0.
Why it matters The repository provides a concrete architecture for isolating agent tasks and managing their access.
Google announced AI-powered defences against scam messages and call-forwarding abuse on Android. The features are rolling out to eligible devices.
Why it matters These protections could help users avoid common messaging and call scams.
Nemotron 3.5 Super is a vision-language model for coding, reasoning, tool use and document understanding, with a context window of up to one million tokens. The release also describes Nemotron 3.5 Bolt, a smaller distilled sibling designed to fit 24 GB of NVIDIA memory.
Why it matters Companies can assess a multimodal model with tool-use support for workflows running through Ollama.
AI repositories, releases and agent tools that gained traction.
Open Scion is an open-source version of Scion, a version-control system built for coding agents. Its listed topics include AI agents, coding agents, Git and Go.
Why it matters Purpose-built version control may help coding agents manage changes.
Muse Spark Open is a multimodal mixture-of-experts model that accepts text, images, audio and video, and generates text. It supports a context window of up to 1,048,576 tokens and is trained for agentic coding tasks such as editing multiple files, debugging and tool use.
Why it matters Teams can assess whether its multimodal inputs and coding-agent capabilities fit their workflows.
New open models from Hugging Face and the papers people discuss.
| Model | Task | Size | Licence | DGX Spark | Popularity |
|---|---|---|---|---|---|
| mistralai/Timestral-28M | time series forecasting | — | other | Runs |
|
| microsoft/UserRepGen-4Bfine-tune of Qwen/Qwen3.5-4B | Text generation | 4.5 B | mit | Runs |
|
| nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | Vision language | 124 BMoE | other | Runs at 4-bit |
|
| nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorrquantized of nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB | Text generation | 0.7 BMoE | other | Runs |
|
| inclusionAI/Ling-3.1-flash | Text generation | 572 B | mit | Too large |
|
The study examines collaboration between pretrained unimodal models on multimodal prediction tasks when some inputs are missing. Using fixed surviving-model anchors, it finds that collaboration can be less accurate than relying on a surviving model alone, and that aggregation can weaken recovery of modality-specific information.
Why it matters Teams should test multimodal systems with missing inputs, as collaboration may reduce reliability.
Researchers derived a 15-slot information schema from professional humanitarian situation reports and used it to assess 121 human and 261 LLM-generated reports. The evaluation covered five disasters and three hazard types, focusing on information practitioners need rather than convenient comparisons.
Why it matters The framework gives teams a practitioner-focused way to assess LLM-generated crisis reports.
Products, money and policy that change how companies use AI.
Apple and Meta are working to make Meta’s Muse AI model run on Mac computers. Meta has shifted from its open-weight Llama models towards proprietary models under Wang, who joined after Meta invested in Scale AI.
Why it matters The collaboration could make a major AI model available on Apple’s computer platform.
Google’s Gemini 4 Argon scored 97.5% on ARC-AGI-3 using a provider adapter harness, but 49.8% with the standard harness. It also scored 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1, with costs listed per task.
Why it matters The results provide benchmarks for comparing a frontier AI model’s capabilities and costs.
Anthropic told investors its quarterly revenue had reached $16.8 billion. Revenue was up nearly 50% as the company prepared for an IPO.
Why it matters The figures signal rapid growth at a major AI company ahead of a potential public offering.
Google is marking Teachers’ Day with AI for Educators in Brazilian Portuguese. Study Mode is designed to help teachers support active, step-by-step learning.
Why it matters The update applies AI tools to classroom teaching and student learning.
Microsoft says general-purpose AI can fall short for legal professionals, who need to understand and review the evidence behind answers. Copilot can bring more of this work together, while decisions remain with legal professionals.
Why it matters The product targets a professional use case where evidence review and human oversight are central.
Timestral offers probabilistic forecasting in four sizes, from 5 million to 209 million parameters. The models are described as CPU-efficient and ready for zero-shot use.
Why it matters The release adds AI forecasting models designed to run efficiently across a range of sizes.
Nvidia, AMD and Dell shares fell as the global technology hardware sell-off deepened on Thursday. The rout was tied to worries about AI funding.
Why it matters Market concerns about AI investment are affecting major hardware companies.
Posts from labs, researchers and builders that people shared most.
NVIDIA introduced Nemotron 3.5 Bolt, which the post says fits into 24GB of memory. It provides no further model specifications.
Why it matters A model that fits into 24GB could be practical for engineers with limited hardware.
GPT-Live-1.1 brings new voices and improved accents to Voice in ChatGPT, Codex and dots. The update is rolling out to users, with API availability coming soon.
Why it matters Teams building voice interfaces can track the rollout and API availability.
NVIDIA’s Nemotron 3.5 Super VL is now available in Perplexity for Pro and Max subscribers. It is also available through the Agent API.
Why it matters Engineers evaluating models for agent workflows can now consider this option through Perplexity and its Agent API.
Sites is now generally available for Enterprise, with US and EU data residency for new Sites in eligible workspaces. Teams can build apps with databases and authentication; the post also demonstrates scheduling updates and using teammates’ connected data with Plugins.
Why it matters The app-building capability combines data residency options with databases, authentication and integrations.
Google DeepMind says SynthID has watermarked more than 180 billion images and videos. The post says the technology helps people distinguish AI-generated content from other media.
Why it matters The scale of watermarking is relevant to organisations considering AI provenance and content verification.
A video shows Gemini 4 Argon acting as a virtual builder in Antigravity to create a multi-tiered pagoda sanctuary. The resulting 3D world is described as fully explorable and built in minutes.
Why it matters The demo shows a generative workflow for creating interactive 3D environments.
CERN and Mistral are working together to explore what large-scale AI models can contribute to particle physics. Their ambition is a single Large Physics Model trained on billions of simulated collisions.
Why it matters The project applies large-scale model development to scientific data and simulation.
A post shares a llama serve command for the GGUF Q2_K version of Muse Spark Open. The author says it runs on two Mac Studios.
Why it matters The example offers a practical deployment signal for local inference with a very large model.
AI Scientist v3 moves from one agent writing a paper to parallel workers that share memory and distilled conclusions. Its evidence-generation process uses isolated agents to run controlled baselines and ablations.
Why it matters The system demonstrates an agent architecture for automating parts of the research loop.
The post warns that prompt injection could expose payment-card details held by an agent. It recommends giving the agent a mock token, allowlisting merchants, setting spending caps and keeping the real card hidden.
Why it matters It gives practical security controls for agents that can make purchases.
Attacks on AI systems, attacks using AI and the defences that work.
Attackers are impersonating popular AI products to distribute malware, steal browser sessions and compromise business accounts. ANY.RUN linked 1,258 public sandbox analyses to these campaigns between April and October 2026.
Why it matters Companies should account for impersonated AI products in efforts to protect business accounts and browser sessions.
Securing AI agents involves more than protecting the information they know. It also requires controlling what they can do.
Why it matters Agent permissions and actions are central security concerns as AI systems gain capabilities.
AI pentesting can uncover vulnerabilities that human testers miss, but it has limitations in consistency, risk assessment and business context. The article compares both approaches and considers where AI agents can improve security testing.
Why it matters Teams should weigh AI’s strengths and limitations when deciding how to test for vulnerabilities.
AI is already being used in digital forensics and incident response, even without a policy governing its use. The article warns that confidence in AI output is growing faster than reliable methods to verify it.
Why it matters Unverified AI findings could undermine security investigations and response decisions.
Jev and other decision models can still be tricked with prompt injection. Avoiding type errors does not mean a model is incapable of other errors.
Why it matters Prompt injection remains a risk for AI decision systems.
The article argues that a guardrail can detect an attack but cannot authorise a refund. It says the term “guardrail” is being used for two different functions.
Why it matters Teams should distinguish between detecting security threats and authorising consequential actions.
An automated pipeline selects and summarises these stories and an AI model translates them. A person does not review each edition. Check the original source before you act on a story.
