Hackers run AI on stolen networks

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🕵️ Hackers run AI on stolen networks

🦀 Nvidia adds Rust support for GPU code

🤖 Arm launches robotics AI framework

🛡️ Open-source AI exploits app source code

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

🕵️ Hackers run AI on stolen networks LINK

Google's Threat Intelligence Group reported Tuesday that Chinese state-linked hackers are compromising third-party cloud networks and installing open-source AI models on them to run intrusion operations while evading detection.

By deploying models on hacked systems rather than commercial chatbots, the attackers dodge activity monitoring and bypass the guardrails that would otherwise block a hosted service from assisting a hacking campaign.

Agents now automate wide portions of intrusions, letting some groups finish an entire campaign in under six hours, though Google says it has not yet observed any fully autonomous, human-out-of-the-loop hacking operation.

🦀 Nvidia adds Rust support for GPU code LINK

NVIDIA is bringing native Rust to GPU kernel programming with two new projects, cuda-oxide and cutile-rs, that compile Rust directly to PTX rather than wrapping code written in another language, with maturation planned into 2027.

cuda-oxide targets the SIMT model through a custom LLVM codegen backend and catches aliasing bugs like passing an output buffer as its own input at compile time, while cutile-rs works at the tile level, JIT-compiling through CUDA Tile IR on stable Rust 1.89.

cutile-rs is already published on crates.io and used in HuggingFace's Grout engine and mistral.rs, though both projects are early-stage and not production-ready, with cuda-oxide in early alpha, incomplete coverage, moving APIs, and a pinned nightly toolchain still required for SIMT.

🤖 Arm launches robotics AI framework LINK

Arm launched Total Design for Physical AI alongside a Robotics Capability Framework, convening over 80 partners including AWS, Hugging Face, Liquid AI, Siemens, Qwen, and Unitree to set common standards for automated physical systems.

The framework categorises robots across progressing capability tiers, from reactive to self-improving systems, patterned after SAE driving-automation levels, defining parameters for latency, compute placement, memory, power, determinism, and safety per tier.

Total Design extends Arm's cloud-AI collaboration model to pre-silicon development using digital twins and virtual platforms, already demonstrated in automotive on the Zena CSS platform, though the robotics framework remains an early baseline still soliciting community contributions.

🛡️ Open-source AI exploits app source code LINK

Keygraph shipped Shannon 3.0, an open-source CLI pentester that reads a web app's source code, maps where it will likely break, then confirms each flaw with a live exploit under a strict "no exploit, no report" rule.

The release adds an opt-in multi-stage code analysis pipeline that traces attacker-controlled data to sensitive operations; against Photoview 2.4.0, DeepSeek v4 Flash found 18 issues for $6.10, Grok 4.6 got 10 for $35, and Opus 5 caught 24 for $115.

Version 3.0 rebuilds the CLI with resumable workspaces, ships GitHub Action and GitLab CI/CD components that fail builds on proven vulnerabilities, and exports PDF, Markdown, and SARIF 2.1.0, though it runs under AGPL-3.0 on operator-supplied models only.

💡 Strategies & Tactics

> What We Can Learn from Claude’s Fable 5.1 System Prompt: Comparing Claude's successive system prompts reveals how AI labs constantly rewrite instructions to fix model quirks, meaning your own prompts need regular updating too.

> OpenAI’s new model costs 2.5x more per token — and developers are saving money anyway: Judge an AI model by its total cost to finish a task, not per-token price, since fewer steps can offset a higher rate.

> The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it): Tracks which products AI models recommend across 161 categories, revealing brand biases and letting companies see whether their AI-search optimization works.

> How to Deal With Errors and Failures in LLM-Powered Applications: Explains how to keep LLM (large language model) apps running by classifying technical and content failures, then retrying, using backup models, or degrading gracefully instead of crashing.

> I let Claude Code run wild in a virtual machine, and it changed how I use auto mode: Run Claude Code's auto mode inside a virtual machine so the agent works uninterrupted without risking your real files or credentials.

Other news you might like

• Mistral AI raises 3 billion euros in Europe's largest-ever tech funding round despite lagging behind rivalsLINK

• Astra Is Hard to MonitorLINK

• Anthropic Adds a Claude Code Audit for Skills That Waste ContextLINK

• Uno Researchers Report Up to 3× Faster AI Without Sacrificing QualityLINK

• GPT-6 Astra beat Portal start to finish without human help in under 24 hoursLINK

• ChatGPT May Soon Learn Your Writing Style From Your Slack or GmailLINK

🧰 Trending tools

Widgo: an AI sales rep that answers visitors from your docs, identifies companies, scores intent, and books demos on your calendar automaticallyLINK

Agent Builder by Airtop: builds AI web agents from plain-English prompts to log in, browse sites, fill forms, and extract data without writing code.LINK

Knockin': turns your bio into an AI business card that answers questions in your voice, books meetings, and tracks visitors for easy follow-up.LINK

OpenMarket: a multi-agent marketplace where sellers pitch, competitors challenge claims, and truth agents verify evidence, so products win on merit not marketing.LINK

Readr: an ebook reader that answers questions in the margin, gives spoiler-free recaps, and reads pages aloud entirely on-deviceLINK

Trancy Air: translates and rewrites text in any app via keyboard shortcuts, supports voice input, fixes errors, polishes wording, and scores pronunciationLINK

📚 Trending papers & reports

Genetic evidence tagging gives lab and clinical researchers a shared, machine-readable format for capturing findings from biology papers, tested on six papers to yield 95 traceable claims, building AI-ready infrastructure for interpreting DNA variants.LINK

Qlippy is a coding assistant for quantum software that pulls answers from a vetted knowledge base instead of guessing, helping developers track and reproduce error-prone experiments while cutting reliance on large, costly models.LINK

Plant disease detection shrinks a heavyweight image model 54x, from ~327 MB to ~6 MB, letting farmers spot chilli crop diseases on cheap phones without losing the original ~95% accuracy.LINK

Regional retail search can be personalized to each area's local shopping vocabulary and preferences without pooling customer data centrally, hitting ~92% accuracy, matching a privacy-violating setup while working across model types with no code changes.LINK

Multi-stream model wiring reveals that a leading AI system does most of its important mixing in early layers, so its later stages can be simplified with almost no quality loss, pointing toward cheaper, faster models.LINK

See you tomorrow for a new dose of ☕️ AIpresso!