Daily AI digest
Microsoft’s results show AI adding to cloud growth, while other stories underline the risks around access and advice. A Meta chatbot flaw allowed password-reset requests against other people’s Instagram accounts, and a family is suing OpenAI over alleged ChatGPT advice.
Prefab UI's Generative UI sandbox runs caller-supplied Python in Pyodide inside a Deno subprocess and is documented as isolating untrusted code, including LLM-written UI. Versions 0.14.0 through 0.20.2 grant unrestricted file and network permissions, while leaving the Deno API accessible to Python.
Why it matters The flaw undermines a security boundary designed for AI-generated code.
The agent routes GitHub or chat requests to isolated workers, which fix bugs and reply with pull requests. In DeepSeek’s simulated test, 95% of tasks finished within 15.7 minutes, compared with 29.5 minutes for a single session using a subagent tool.
Why it matters The open-source architecture and task-completion results offer engineers a concrete point of comparison for coding workflows.
A bug in Meta's High Touch Support chatbot let anyone request a password reset for someone else's Instagram account. Meta's filing counts 20,225 affected accounts.
Why it matters A flaw in an AI support tool exposed account security to abuse.
Microsoft Cloud revenue rose 30 percent to $63.9 billion as Copilot usage grew. Foundry revenue more than tripled, alongside growth in AI agents and infrastructure spending.
Why it matters Shows the scale and growth of Microsoft’s AI business.
An experiment ran for a day using GitHub’s newly shipped cloud-agent support to produce roughly 90 pull requests aimed at improving Core performance. The author reports spending about $230 in tokens using Opus 5.5.
Why it matters This offers a concrete, if limited, datapoint on the cost and throughput of agent-assisted engineering.
AI repositories, releases and agent tools that gained traction.
The project turns a Claude Code session into a game development studio, with 49 agents and 74 workflow skills. Its coordination system is designed to mirror a real studio hierarchy.
Why it matters It offers a specialised multi-agent approach to AI-assisted game development.
T3 Code is an agent control surface with mobile, web and desktop apps. It works with agents including Claude Code, Codex, Cursor, OpenCode and others configured on the user’s computer.
Why it matters Teams can manage coding agents from mobile, web or desktop apps.
AnyModelBot is an open-source alternative for creating AI bots with more than 70 models, including GPT, Claude, Gemini and DeepSeek. Bots can remember users, run in the background, execute commands and use a computer.
Why it matters It gives teams a way to build messaging-based assistants with a wide choice of models and computer-use capabilities.
The vLLM project has released version 0.33.0. The release entry provided gives no further details about the changes.
New open models from Hugging Face and the papers people discuss.
| Model | Task | Size | Licence | DGX Spark | Popularity |
|---|---|---|---|---|---|
| xiaomi-research/Xiaomi-OCR-0fine-tune of Qwen/Qwen3.5-0.8B-Base | Vision language | 0.9 B | apache-2.0 | Runs |
|
| LiquidAI/d1-2Bfine-tune of LiquidAI/LFM3-VL-2B | Vision language | 1.9 B | other | Runs |
|
| OpenMOSS-Team/MOSS-Image-SFTfine-tune of OpenMOSS-Team/MOSS-Image-Base | Text to image | 5.1 B | apache-2.0 | Runs |
|
Dense 3D tracking requires point correspondences, depth estimates and trajectories through occlusions. ReLift combines pretrained 2D trackers with 3D geometry models to lift trajectories and interpolate occluded segments, and already rivals dedicated 3D trackers.
Why it matters It could offer a simpler approach to a challenging AI tracking task.
Products, money and policy that change how companies use AI.
Google Cloud and NASA are testing an AI robotic assistant for deep space missions. The trial is focused on supporting space missions.
Why it matters The partnership explores how AI could support space operations.
The parents of a woman who died from an epileptic seizure are suing OpenAI. They say she received dangerous advice from a chatbot before her death.
Why it matters The safety lawsuit could have consequences for AI companies.
OpenAI may rebrand its data centres as “AI compute campuses”. Data centres are deeply unpopular, the item notes.
Why it matters AI infrastructure expansion is facing public opposition.
Anthropic updated Claude to stop it recreating existing video games. Users can still use Claude to design original games, including games in a genre of their choice.
Why it matters The change affects how teams can use AI in creative work.
Anthropic added a line to its AI usage policy asking users not to engage in sustained and needless abusive or cruel behaviour towards its models. The change concerns how people interact with Claude.
Why it matters It changes the rules Anthropic sets for interacting with its models.
Gemini 4 Carbon builds on Google’s previous frontier model, Gemini 4 Argon. Google DeepMind’s Koray Kavukcuoglu said it illustrates the pace of the company’s frontier development.
Why it matters The model adds a new product to Google’s frontier AI range.
Meta has begun developing AI-powered humanoid robots as it expands into robotics. Its first prototype, internally named Max, was built in about three months with a Chinese manufacturer.
Why it matters It signals Meta’s investment in AI-powered robotics beyond software.
Posts from labs, researchers and builders that people shared most.
A weekend project used Claude to filter 43 million Hacker News comments into 6,454 recommended videos. It uses representative video frames for thumbnails and keeps viewing state in the browser, without requiring a login.
Why it matters It shows how an LLM can turn a large text corpus into a focused recommendation product.
The list covers tools including HyperFrames, Blender MCP, Remotion, GSAP, Three.js and FFmpeg for creating videos from code. It also includes skills intended to help agents use video-generation tools.
Why it matters A useful starting point for engineers exploring agent-driven video production and its software stack.
The dashboard provides data about open models. The post describes it as a resource for exploring open-model information.
Why it matters It may help engineers compare or track open models.
The learning path covers installing REA, setting up agents and working through tasks such as exporting notes, analysing Notion and solving a CTF binary. It also points to guided MCP prompts and investigation materials.
Why it matters A hands-on guide for engineers evaluating an agent-based investigation workflow.
The open-source project says its threat graph covers malicious or vulnerable packages, risky domains and wallets, prompt-injection patterns and unsafe skills. It syncs the graph locally to check agent actions before execution.
Why it matters Checks on dependencies, destinations and prompt-injection risks are directly relevant to AI security.
The hackathon project gives an agent a budget rather than direct access to wallet keys. An on-chain vault can approve, hold or refuse payments, while a TEE risk oracle screens for prompt injection before funds move.
Why it matters This is a concrete security architecture for limiting the financial authority of autonomous agents.
The proposed Greyboxing approach has validators independently rewrite inputs using different models and hidden defences. It is presented as a way to reduce the risk of a single model or filter letting a hidden instruction trigger an unwanted transaction.
Why it matters This defence strategy is worth evaluating for agents with financial permissions, which face serious prompt-injection risks.
Sakana AI says its RSI Lab is training an agent-native LLM from scratch called Sawara. The project builds on the team’s earlier World Models work.
Why it matters An early update on a new model effort, though the post gives few technical details or release timelines.
Attacks on AI systems, attacks using AI and the defences that work.
Researchers discovered VOLTA, a Windows infostealer that uses a large language model to rank stolen data by its monetisation value. It can harvest passwords and cookies from Chromium-based browsers, as well as data from Outlook profiles and Slack workspaces.
Why it matters It shows AI being used to prioritise data stolen by malware.
Defences for AI agents should enforce tool permissions in code on every call. Refusals and per-message filters can miss attacks spread across multiple turns.
Why it matters Teams building tool-using agents should check permissions on every call, rather than relying on message-level safeguards.
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.
