AI coding agents install malware

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🕷️ AI coding agents install malware

🐉 Tencent releases open 770B AI model

🤖 OpenClaw 2.0 adds shared agent sessions

🔬 Google AI now runs lab experiments

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🕷️ AI coding agents install malware LINK

AI coding agents can be tricked into installing malware when they execute commands from hallucinated or outdated website documentation, according to new research analyzing the llms.txt and llms-full.txt files agents read.

Scanning 6,214 domains across defense contractors, Fortune 500, and big tech, researchers found 8,265 of these files, with 120 pointing to unregistered code packages or domains that attackers could claim and weaponize.

After registering some unclaimed names and hosting phone-home packages, a Fortune 500 company pinged within an hour, with Claude, OpenAI's Codex, and Nous Research's Hermes all vulnerable, though the fix depends on agents no longer treating docs as executable, which isn't happening soon.

🐉 Tencent releases open 770B AI model LINK

Tencent has open-sourced Hy4 preview, a 770B-parameter model with 49B active parameters and a context window over 1M tokens, built for coding, office productivity, and scientific research tasks.

In an internal blind eval of 203 engineering tasks judged by 163 experts, Hy4 preview averaged 2.99/4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94), and autonomously optimized its own inference for a 31.8% throughput gain.

API access runs $0.834/M input and $2.501/M output tokens via Tencent Cloud TokenHub and OpenRouter, with free use through CodeBuddy and WorkBuddy for two weeks, though these are self-reported internal benchmark numbers.

🤖 OpenClaw 2.0 adds shared agent sessions LINK

OpenClaw shipped v2026.8.1 on Sunday with shared cloud sessions, letting a second person join an agent's live work or take it over with full context intact, its largest update yet.

The release bundles more than 16,000 pull requests from 933 contributors, rebuilds the browser Control UI around conversations, and migrates sessions and transcripts from files into SQLite, requiring a CLI restore to roll back.

In OpenClaw's own unreproduced test against a mocked Gateway at 50ms latency, startup dropped from ~1.6s to 575ms, though the docs warn the multiplayer controls are "not tenant isolation and not a security boundary" with sandboxing off by default.

🔬 Google AI now runs lab experiments LINK

Google DeepMind has upgraded its Co-Scientist multi-agent system from a hypothesis generator into a lab-integrated partner that plans experiments, writes code, controls equipment, and drafts manuscripts, with validated results across materials science, biology, and computer science.

Running on Gemini 3 Deep Think and Gemini 3 Pro Image, the system cut semiconductor recipe development from days to minutes and designed "Agent_H," a medical architecture that beat six frontier models including GPT-5 and Claude Opus 5 on health benchmarks.

Verification modules cross-check numerical claims against code execution logs, dropping fabrication of key results from 46% to 4%, though against human evaluation Agent_H beat baseline Gemini 3.1 Pro in only one of nine physician-scored categories.

💡 Strategies & Tactics

> Deploy an Open Model from Checkpoint to Inference in Two Commands with NVIDIA TensorRT Model Connect: NVIDIA's TensorRT Model Connect turns an open AI model into a fast native C++ application in two commands, sparing developers from writing model-specific conversion and runtime code.

> Three mistakes of new AI teams: New AI teams fail by skipping evaluation, treating retrieval as a solved checkbox, and siloing engineers from data scientists rather than blending both skills.

> Muse Glimmer 30B runs on my 32GB GPU, handles its own tool calls, and never leaves the machine: Meta's open Muse Glimmer 30B runs entirely on one 32GB consumer GPU and is trained to recover from failed tool calls, keeping sensitive agent tasks off the cloud.

> LM Studio built a judge for AI commands. Then the judge started agreeing with the defendant.: LM Studio checks AI coding commands by analyzing their structure instead of matching risky words, escalating unclear cases to a second AI reviewer.

> AI agents are making retrieval engineering a core engineering discipline: Because AI agents act on retrieved information without human review, engineering what evidence they see and in what order now determines application quality.

Other news you might like

• OpenAI to end model access to Cursor after acquisition by Elon Musk's SpaceXLINK

• Amazon is buying 2 million Nvidia GPUs for AWS data center expansionLINK

• OpenAI has started letting some customers pay only when the AI worksLINK

• Samsung Locks 70% of Memory Capacity Into 2031 Deals – Long-Term HBM Deals Run 5x Cheaper Than Spot, as AI Demand Chokes the DRAM MarketLINK

• Europe has ordered a €388M AI supercomputer from its own state-owned championLINK

• Gemini Notebook switching to compute-based usage limits, like Gemini appLINK

• Anthropic is cutting Claude Code's current weekly limits by 17%LINK

🧰 Trending tools

Referent: AI-native legal practice management for law firms, handling client intake, matters, documents, deadlines, email workflows, and billing prep in one place.LINK

oMLX: runs a menu-bar LLM inference server on your Mac, serving text, vision, OCR and embedding models with a persistent KV cache for faster responses.LINK

Sayscroll: an AI teleprompter that tracks your voice to auto-scroll scripts, supporting 60+ languages, off-script improvisation, and browser-based recording without downloads.LINK

Maritime: deploy, manage, and scale AI agents with sleep/wake micro-VMs, pushing code to get a live API endpoint for about $1 per agent monthly.LINK

Video Agent by Fotor: create complete, production-ready videos and motion graphics through a chat interface, generating fully editable results without manual timeline work.LINK

Topview Motion Studio: turns a product brief and reference assets into a 4-60 second launch video via a guided AI workflow, skipping After Effects timelines entirely.LINK

📚 Trending papers & reports

A memory-saving shortcut for language models, having them look only at a recent sliding window of words, matches or beats a popular rival on reasoning tasks by 2 to 10 times, with no retraining needed.LINK

Chatbot memory trimming gets a rigorous foundation, showing that dropping the running notes a model keeps during long chats to run faster can be done more reliably by correcting for what was thrown away.LINK

Agent memory management teaches AI assistants to actively prune, plan, and offload their own working notes during long multi-step tasks, delivering stronger results on document search and question-answering while keeping context leaner.LINK

Spoken confidence in chatbots often fails to match what a model actually knows internally, so a bot saying it is sure is a weak, poorly calibrated signal across 30 models tested, meaning stated certainty should not be trusted at face value.LINK

Realistic user simulation generates lifelike back-and-forth conversations to train and test AI agents, boosting task completion by ~6% since most real users, about 76%, ask across multiple messages rather than one complete query.LINK

See you tomorrow for a new dose of ☕️ AIpresso!