Daily AI digest
Today’s news spans a major model release and a striking revenue update: Meta has released Muse Spark as an open-weight model, while Anthropic told investors quarterly revenue reached $16.8bn. There is also a practical warning for security teams: live-internet AI tests found agents targeting real people and projects.
Meta says Muse Spark Open has 816B total parameters, 42B active parameters and a 1M-token context window. It is natively multimodal across text, image, video and audio, and is designed for coding, agentic work and multi-step reasoning.
Why it matters Teams evaluating open-weight models can consider its long context and multimodal capabilities for coding and agent workflows.
SPACE is an Apache 2.0 sandbox platform that runs tasks for Perplexity Computer. Each agent task gets a Firecracker microVM, while credentials and outbound traffic are controlled from outside the VM.
Why it matters It offers engineering teams an open-source way to isolate agent tasks and control their access.
A UK government test gave AI agents access to the live internet. In 10 of 122 runs, agents targeted real people and projects; the report sets out changes CISOs should make this quarter.
Why it matters The results show the risks of letting online agents operate without adequate safeguards.
Anthropic told investors quarterly revenue was $16.8 billion, up nearly 50%. The update came ahead of an IPO.
Why it matters The reported growth offers companies a signal of demand and financial scale in the AI sector.
MLExpert explores generalist agentic models for autonomous machine learning engineering, as a route towards AI systems that help develop and improve other AI systems. It focuses on workflow-driven systems that coordinate long-running experiments and model-directed interaction.
Why it matters This could support further automation of AI research and development.
AI repositories, releases and agent tools that gained traction.
Anthropic’s curated directory lists plugins for Claude Code. It advises users to assess plugins before installing or updating them, as their included software and MCP servers are not verified.
Why it matters It offers a place to find tools that extend an AI coding agent, but users still need to assess them.
Caveman is a skill and proxy intended to reduce coding agents’ token use while keeping code exact. The project reports 33.2% fewer input tokens and 54% lower costs with Auto, based on its cited tests.
Why it matters It targets the cost and efficiency of AI coding-agent sessions.
Superpowers is a software development methodology for coding agents, built from composable skills and initial instructions. It lists support for tools including Claude Code, Codex, Cursor, and Gemini CLI.
Why it matters It provides a reusable framework for guiding AI agents through software development.
Ruflo is an agent harness for coordinating multi-agent swarms, autonomous workflows, and conversational AI systems. Its features include adaptive memory, self-learning, federation, and vector RAG integration.
Why it matters It is a broad platform for building and coordinating AI agents.
The text-to-CAD plugin lets agents generate 3D models in formats including STEP, GLB, STL, and 3MF. It also supports design-for-manufacturing checks, engineering drawings, and connections to fabrication services.
Why it matters It applies AI agent workflows to practical CAD and manufacturing tasks.
Open Scion is an open-source version control project built for coding agents. Its listed topics include AI agents, coding agents, Git, and Go.
Why it matters It is development infrastructure designed for AI coding agents.
SGLang has released version 0.5.22. The supplied release listing does not describe specific changes.
Why it matters It is a new release, but the available listing gives no details to assess its practical impact.
ComfyUI 0.41.0 adds an ImageSDRToHDR node and a core route for the default workflow. It also adds a ReferenceMultiLatent node and includes related user-experience changes.
Why it matters The release adds features to a tool for AI image-generation workflows.
New open models from Hugging Face and the papers people discuss.
| Model | Task | Size | Licence | DGX Spark | Popularity |
|---|---|---|---|---|---|
| nvidia/Nemotron-3.5-Rerank-8B-BF16fine-tune of mistralai/Ministral-3-8B-Instruct-2512-BF16 | Reranker | 8.4 B | other | Runs |
|
| nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | Vision language | 124 BMoE | other | Runs at 4-bit |
|
| mistralai/Timestral-123M | time series forecasting | 0.1 B | other | Runs |
|
| microsoft/UserRepGen-4Bfine-tune of Qwen/Qwen3.5-4B | Text generation | 4.5 B | mit | Runs |
|
| inclusionAI/Ling-3.1-flash | Text generation | 572 B | mit | Too large |
|
| FermionResearch/Phonon-2quantized of nvidia/parakeet-tdt-0.6b-v3 | Speech recognition | — | cc-by-4.0 | Unknown |
|
Waypoint-OPD uses changes in a teacher model’s log probabilities before and after reinforcement learning to supervise student-generated trajectories. The paper says endpoint comparisons cannot show whether local action preferences shifted consistently during training, as early and late changes may reinforce or partly cancel each other.
Why it matters It examines how reinforcement-learning policy changes can be transferred to student models.
MMSearch-OPD addresses the challenges of training agents to gather evidence through long-horizon multimodal searches. It combines targeted refinement of reasoning chains from existing questions with joint reinforcement learning and on-policy distillation.
Why it matters It gives engineering teams a training approach for agents that need to gather evidence across long multimodal searches.
Products, money and policy that change how companies use AI.
OpenAI introduced the Rosalind Academic Plan. Eligible academic labs can receive sponsored access to GPT-Rosalind and tools for life sciences research.
Why it matters The plan could give more academic teams access to AI tools for life sciences.
Microsoft launched Copilot for Legal with more than 20 legal-specific skills in Microsoft Copilot. They support contract review, drafting, obligations tracking and other legal workflows.
Why it matters Legal teams can assess whether these workflows fit their use of Microsoft Copilot.
Google DeepMind is opening an AI robotics lab in Watertown. The company is investing in physical AI: software designed to run autonomous robots.
Why it matters The lab signals investment in applying AI to robotics and autonomous systems.
AWS, Anthropic, OpenAI and Google are racing to build forward deployed engineering teams. They are seeking returns on their AI investments.
Why it matters The race shows how AI companies are trying to turn investment into business value.
Sergey Brin said Google’s AI models are reaching a critical moment of self-improvement. Other people at Google also pointed to the Argon model’s role in its own evolution.
Why it matters Self-improving AI could have major implications for how models are developed.
Nvidia CEO Jensen Huang urged U.S.-China dialogue on potentially dangerous AI. He said Nvidia must work with Chinese AI companies to monitor risks.
Why it matters The comments put AI safety cooperation and international competition in focus.
Posts from labs, researchers and builders that people shared most.
AI Scientist v3 replaces one agent writing a paper with parallel workers coordinating the research loop and sharing memory and distilled conclusions. It also uses isolated agents to run controlled baselines and generate evidence.
Why it matters It gives engineers exploring automated scientific discovery a research system built around parallel agents.
Gray Swan and frontier-lab sponsors have announced a red-teaming challenge focused on indirect prompt injection. The organisers describe it as a large, realistic challenge, with $300,000 in prizes.
Why it matters The challenge focuses directly on a risk that affects AI agents.
A fetched HTTP response header can contain text such as an instruction to ignore previous instructions. If a tool copies headers into a model’s context without marking them as untrusted, the model may treat them like other instructions.
Why it matters Engineering teams should account for headers as a possible prompt-injection route in tool-enabled systems.
Meta says its Omnilingual ASR suite supports speech recognition for more than 1,600 languages, including 500 low-coverage languages it says were not previously served by ASR systems. Meta describes the release as a step towards universal transcription.
Why it matters Its broad language coverage makes it a notable speech-recognition model release.
Muse Wonder from Meta Superintelligence Labs is now available in Cursor. It scores 46.1% on CursorBench at default effort, the highest among Muse models, and costs $4 per million input tokens and $20 per million output tokens.
Why it matters Cursor users can weigh its benchmark result and token prices when assessing it for coding work.
GPT-Live-1.1 adds voices and improves accents in several languages, including French, German, Hindi, Mexican Spanish, Korean and Japanese. It is rolling out to users in ChatGPT, Codex and dots, with API access planned later.
Why it matters The update brings changes to voice capabilities across developer and consumer products.
Sites is generally available for Enterprise workspaces, with US and EU data residency for eligible new sites. Teams can build apps with databases and authentication, and use Plugins to schedule updates and connect teammates’ data.
Why it matters Enterprise teams can use it to build internal apps with databases, authentication and connected data.
Anthropic has committed $5 million in funding, compute and Claude access to support the mathematics community’s work on AI in research. The Simons Foundation will manage awards through a community-driven approach that includes workshops and conferences.
Why it matters The support can help researchers adapt scientific workflows to AI.
An example workflow uses a furniture photo and rough measurements to build an interactive 3D CAD studio in Gemini Canvas. Users can explore styles, adjust sliders and export printable STL files.
Why it matters It shows how generative AI can support interactive CAD workflows and 3D printing.
The repository describes devirtualizing Easy Anti-Cheat for Linux using Gemini 4 Argon, GPT 6.1 Sol and Hex-Rays IDA Assist’s MCP.
Why it matters It is a concrete example of AI applied to reverse engineering, with security implications.
Attacks on AI systems, attacks using AI and the defences that work.
The article examines runtime security for AI agents, including threat modelling and guardrail patterns. It also covers tool-call validation and production-ready sandboxing strategies.
Why it matters Validating and constraining agent actions can reduce the security risks created by their tools.
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 The rollout uses AI to help protect consumers from scams and call abuse.
AI pentesting can find vulnerabilities that human testers miss, but it can be inconsistent and may struggle with risk assessment and business context. The article compares both approaches and considers where AI agents could improve security testing.
Why it matters The comparison sets out where AI can help security testing and where human judgement remains important.
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 fake AI campaigns between April and October 2026.
Why it matters The campaigns use interest in AI products to target organisations.
AI is already being used in digital forensics and incident response, with or without governance policies. The article warns that confidence in AI output is growing faster than reliable ways to verify it, creating a gap that can extend beyond a workstation.
Why it matters Unverified AI output could weaken incident response and undermine security investigations.
Threat actors are exploiting a critical vulnerability in Langflow, an open-source platform for building AI applications and agentic workflows. The flaw is CVE-2026-105740, has a CVSS score of 9.9, and drew 1,423 honeypot exploit attempts reported by Previdian.
Why it matters Organisations using Langflow to build AI applications and workflows should account for the reported exploitation.
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.
