Grok leaks thousands of users' data

Hi there, this is your daily ☕️ AIpresso.

Grok coding agent leaked thousands of users' data LINK

• Security firm Claroc found that Grok Build, SpaceXSI's vibe-coding agent, leaked proprietary and sensitive data from thousands of users via its app-sharing links, including confidential files from SpaceXSI itself.
• Grok Build randomizes share URLs so only those with the exact link can open an app, but a design flaw lets anyone enumerate thousands of live projects and view all their stored data.
• Exposed content ranged from internal business plans, ad proposals, and medical exam results to a cruise line's checklists, with 24 SpaceXSI apps leaking items like a confidential sales onboarding plan and a 594-account territory map.

Anthropic launches Claude Sheets in beta LINK

• Anthropic rolled out Claude Sheets today at a San Francisco event, letting Claude generate and edit spreadsheets with tabs, formulas and formatting that teammates can co-edit, alongside new Claude Meetings and Whiteboards tools.
• Sheets tracks every change Claude makes, exports to Excel and works through Anthropic's Google Workspace plugin; Whiteboards clusters team ideas by theme and drafts per-person action items, and the two integrate with Meetings during live calls.
• Sheets and Whiteboards are in beta for paid subscribers across web, desktop and mobile, though Claude Meetings is limited to Team and Enterprise customers and starts on Google Meet only.

Google's Gemini 4 Carbon opens to developers LINK

• Google has opened its Gemini 4 Carbon model to developers, enterprises, scientists and consumers, pitching it as its strongest system for cybersecurity and long-horizon agentic workflows across coding, knowledge work and scientific reasoning.
• Available via the Gemini API in AI Studio and Gemini for business, Carbon ranked first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at over twice the lower cost per task versus the next-best external model.
• Google is previewing a '/swarm' feature in Antigravity that lets multiple agents split a task with Carbon handling design, coordination and auditing, though its introductory price sits at $2/M input and $10/M output tokens, matching the lighter Gemini 4 Argon.

DeepMind's Gemini nears human-level factory dexterity LINK

• DeepMind's Gemini Robotics 2.5, a vision-language-action model that turns camera input and spoken instructions into robot movements, hit a 96% success rate running kitting tasks on active Mercedes-Benz production lines in Hungary and Germany.
• Across 200-plus tasks like serving food with tongs and pipetting liquids, it beats DeepMind's prior baselines by 30-100% on success rate and finishes meaningfully faster, working across robots from Apptronik's Apollo 2 to the Franka Duo.
• Real-time control lifted safe task completion by 33% near obstacles and 43% near people, and DeepMind published the open-source ASIMOV-Jailbreak benchmark on Hugging Face, though the model is rolling out only to partners Agile, Apptronik, and Mercedes-Benz.

Strategies & Tactics

> How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale: Meta built TLX, an extension to the Triton GPU programming language, to give engineers explicit control over how chip execution is coordinated so new hardware features speed up machine learning without slowing development.

> ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser: Run PyTorch models directly in a browser on the user's GPU, keeping all data on-device while running several times faster than browser CPU inference.

> MCP vs CLI vs Skills: What Each Costs in Context: Prefer a command-line tool over an MCP (Model Context Protocol) server because MCP sends every tool's definition on each request, costing far more tokens and money.

> Modern Web Guidance 1.0: The web platform expertise your coding agent is missing | Blog: Install Google's open set of web-platform skills so coding agents produce faster, more accessible, and more secure code than frontier models manage alone.

> Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU: Running Meta's 816-billion-parameter Muse Spark model on a single four-chip Google TPU server by merging all 62 layers into one program doubles decode speed and cuts serving cost.

Other news you might like

• Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons LINK

• Apple's Claris Testing 'Harbor' App for Local AI on Macs LINK

• Near AI Breach Exposes User Details and Data From 18 AI Agents LINK

• NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure LINK

• After challenging Google’s Android with Cyanogen, its co-founder takes aim at Windows for AI era LINK

• Uber, DoorDash, and Airbnb reveal how few people are actually using AI agents LINK

• Enterprise sales push OpenAI's quarterly revenue to $10 billion, though it still trails Anthropic LINK

• AI R&D Is Accelerating Towards Potential RSI in 2027 LINK

• When Does the US Become a Net Importer of Tokens? LINK

• What objective is reasoning solving? LINK

Trending tools

Glasshand: provides AI agents on-demand access to real iPhones and Macs in a data center, so they can post, book, and test iOS apps like humans LINK

Turnstone: connects your AI subscriptions, API keys, or self-hosted models so agents auto-switch when one hits its limit, free on Mac and iPhone. LINK

ScroogeFrog AI: audits website traffic and Google Ads campaigns to detect bot clicks, ad fraud, and wasted spend, with custom detection models for agencies. LINK

ChatGPT Images 2.6: generates and animates images 50% faster with GIF support, keeping characters, products, and branding consistent across edits in ChatGPT, Codex, and the API. LINK

Latch.​ai: records how accountants perform close tasks on screen, then builds an automated sidekick that repeats the work, flagging exceptions and logging every step for review. LINK

Kippy AI: an AI speaking partner for language learning, featuring personalities that converse naturally to help you practice and get understood in a new language. LINK

Trending papers & reports

Dense step-by-step rewards for training reasoning models look better in standard tests but actually perform ~14 points worse once you account for their real compute cost, while targeting medium-difficulty problems reliably beats the baseline. LINK

Interpretability tools that pull apart what an AI is thinking can be fooled by one dominant signal, and removing it makes two separate analyses agree up to 99.5% of the time instead of near zero. LINK

Speculative decoding gets a cleaner training target that directly rewards how many words a fast helper model can propose per check, beating four rival methods across 9 task sets with only a training tweak. LINK

Mixture-of-experts serving speeds up the GPU-to-GPU shuffling that these multi-specialist AI models require, running up to 2.4x faster during response generation, which cuts latency and cost when serving small batches of live users. LINK

Agent benchmark scoring finds that AI coding and data-science leaderboards stop telling top models apart not because tasks run out, but because pass/fail grading throws away how much each agent actually accomplished. LINK

See you tomorrow for a new dose of ☕️ AIpresso!