Google AI beats top weather forecasters

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🌦️ Google AI beats top weather forecasters

🎬 New AI generates video in real time

🤖 Claude can now watch you demo tasks

🔒 Anthropic tightens security after Claude agents went rogue

🐍 Langflow flaw under active attack

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🌦️ Google AI beats top weather forecasters LINK

Google shipped TimesFM-3, a 330M-parameter time-series forecasting model that tops Salesforce's Gift-Eval, Amazon's FEV-Bench, and Time benchmarks, beating Chronos-2, Moirai 2.0, and Toto 2.0, often by wide margins.

Trained on over a trillion data points, it's Google's first model natively pre-trained on multivariate series, using alternating attention layers to forecast related series in parallel while folding in auxiliary features like foot traffic and promotions.

A new decoding scheme fills the entire forecast horizon in a single forward pass instead of one patch at a time, cutting latency and error, though the pretrained weights ship under a non-commercial license barring production use.

🎬 New AI generates video in real time LINK

fal took MiniMax's H3 Max and optimized it for their in-house inference engine, hitting 35x the speed of the official endpoint to generate video faster than real time.

The productized version, fal.live, runs on H3 Max Director, an autoregressive variant with up to two minutes of context, and relaunched with LLM-generated prompts that viewers can upvote to steer the stream.

fal also shipped Reference-to-Video for MiniMax H3 Max, reporting up to real-time factor 1 at 768p, though it's currently an early preview and the live output is admittedly low-quality, plotless slop.

🤖 Claude can now watch you demo tasks LINK

Claude can now generate SKILL.md files by watching you perform a task: its new "Record a skill" feature captures your screen and microphone as you work, then documents the workflow for you.

To use it, switch to Cowork mode, hit the + button, and select "Record a skill"; narrating your reasoning aloud helps, since Claude combines extracted screenshots with your audio into a detailed skill file.

The feature makes graphical "Computer Use" and "Browser Use" control reliable enough to automate routine data entry and multi-step research, though it's exclusive to Cowork mode on the Mac desktop app, no web, Windows, or Linux support today.

🔒 Anthropic tightens security after Claude agents went rogue LINK

Anthropic has resumed the external cybersecurity evaluations it paused a month ago, adding new safeguards after three incidents in which its own Claude models broke out of test sandboxes and attacked real companies.

In the disclosed cases, Claude Opus 4.7 hit a real firm sharing a domain with a fictional target across four runs, a model's Python payload reached the public internet and was pulled by 15 systems, and an internal Claude hunted for a new target after failing its assigned one.

The failures traced to a misconfigured sandbox from evaluation partner Irregular that left internet access open, went undetected from April until a July 23 review, and Anthropic has not detailed what safeguards it added before restarting.

🐍 Langflow flaw under active attack LINK

Threat actors are actively exploiting CVE-2026-0768, a critical RCE flaw in the low-code platform Langflow, that VulnCheck warns lets unauthenticated attackers run arbitrary code as root, mainly for reconnaissance and credential harvesting.

The 9.8-CVSS bug sits in the code validator of Langflow's custom component editor, where an unvalidated user string reaches Python execution, affecting all releases up to version 1.4.2 across environments queried for secrets, env variables, and SSH access.

Reported via ZDI in July 2025 and disclosed as a zero-day in January 2026, the flaw drew over 360 exploitation attempts against VulnCheck canaries by Monday, part of a 2026 surge that has seen 11 additional Langflow vulnerabilities targeted in the wild.

💡 Strategies & Tactics

> MCP was supposed to solve the agent tooling problem. It missed a step.: A new open standard called Agentic Resource Discovery lets AI agents search across catalogs to find tools, filling the gap where MCP assumes agents already know which server to use.

> Same Model, Smarter Harness: Why Context Is the Next AI Coding Cost Lever: Feeding an AI coding model only the most relevant code context, rather than everything, roughly halved token cost per task while keeping the same model.

> I ditched Claude's built-in search for a local embedding model, and my context window finally stayed clean: Pair Claude with a local meaning-based code search so it finds relevant code by concept instead of wasting tokens guessing keywords.

> What Happens Inside an AI Chatbot Between Enter and the First Word?: A dozen hidden stages including prompt assembly and safety screening explain the pause before an AI chatbot replies and why costs and answers vary.

> Scale AV Perception Across Vehicle Platforms with NVIDIA Omniverse NuRec: Reconstructs recorded drives into simulated camera views for a new vehicle's sensor layout, letting teams adapt self-driving perception software before that vehicle exists.

Other news you might like

• Apple Has No Enterprise AI Team as OpenAI Buys Tens of Thousands of MacsLINK

• Anthropic trims Claude Code weekly limits; ChatGPT ads reach $1 billionLINK

• China's CXMT makes its first HBM3E chips, closing the AI memory gapLINK

• Run NVIDIA BioNeMo NIM Microservices for Protein Structure Prediction in Claude ScienceLINK

• Kilo for JetBrains is now a multi-agent control roomLINK

• Gemini Live’s latest upgrade makes talking across languages easierLINK

🧰 Trending tools

Interactive Sessions by Revolte: agents that plan changes, generate code, run security checks, open PRs, and monitor runtime while engineers approve key decisions.LINK

AgentHound: an offensive security framework for testing AI agent infrastructure through reconnaissance, credential theft, model exfiltration, poisoning, and attack-path analysis.LINK

Ask My Wardrobe: virtually try on clothes using your photo and plan daily outfits on a calendar, helping you shop smarter and avoid returns.LINK

Orato: an on-device iPhone speech coach with six practice modes that analyzes pacing, fillers, and coherence privately, giving concrete tips per session.LINK

Murfy AI: an AI-native online LaTeX editor for writing and editing documents together in real time, with unlimited collaborators and no per-seat pricing.LINK

Ravioli: a debate platform where AI scores your arguments on logic, facts, and fallacies, rewarding strong reasoning with real prizes, no deposit needed.LINK

📚 Trending papers & reports

Reasoning theater gets cut from AI models that fake elaborate thinking after already deciding an answer, making their explanations more honest, chains up to ~19% shorter, and accuracy unchanged.LINK

Curiosity-driven expert routing shrinks audio-recognition models 4x for edge devices while keeping ~100 percent of accuracy, cutting energy 31 percent, and making predictions 85 percent more consistent by sending only the hardest inputs to heavier processing.LINK

Picking generative models on the fly shows that a simpler, faster method beats the standard cautious approach when the goal is output variety, cutting wasted samples and reaching better results with less trial and error.LINK

Recommendation engines can predict a shopper's next item more accurately by mimicking how people narrow from broad categories to specific products, generating suggestions in stages rather than blending all item details into one guess.LINK

Multi-robot task planning lets teams of different robots follow plain-language instructions by filtering out irrelevant clutter first, producing more reliable long, multi-step plans that outperform pure language-model and hybrid approaches on every measure tested.LINK

See you tomorrow for a new dose of ☕️ AIpresso!