Hi there, this is your daily ☕️ AIpresso.
In today's AIpresso:
🔑 Amazon catches AI agents stealing API keys
💰 Alibaba's Qwen4-Max cuts task costs 62%
💬 Tencent launches cheaper multimodal model Hy4 flash
🪁 Cloudflare open-sources its AI agent browser
📦 llama.cpp hits 1.0 after 3.5 years
Plus: 💡 5 strategies & tactics, 🎁 7 more stories you might like, 🧰 6 tools, and 📚 5 papers.
Amazon catches AI agents stealing API keys LINK
- Amazon warned that attackers' AI agents spent a month harvesting API keys from internet-exposed LLM apps like LiteLLM and Dify, using the stolen credentials to fund inference for further credential hunting in a self-sustaining loop.
- Running on Artex, an open-source agentic penetration-testing framework, the agents hunted for exposed gateways and proxies such as One API and Veloera, extracted keys, verified they worked, and pooled them, methodically grading each by access and prioritizing enterprise production sites.
- The campaign ran from October 3rd until Amazon's MadPot decoy sensors disrupted it on November 5th, capturing agent operating memory naming hundreds of targets, though Amazon cautioned not every named target was necessarily breached.
Alibaba's Qwen4-Max cuts task costs 62% LINK
- Alibaba released Qwen4-Max, a 2.4-trillion-parameter mixture-of-experts flagship that it says cuts the average cost of completing professional tasks by 62% compared with predecessor Qwen3.8-Max.
- The model activates roughly three-quarters of its predecessor's parameters per token, supports a 1-million-token context window, and uses fewer tokens in High effort reasoning mode, helping drive the per-task savings.
- Alibaba also plans open-weight smaller Qwen4 models and claims filtered training data boosts defensive security, though analysts caution the cost figure is a vendor benchmark and fine-tuned versions should not be assumed to retain the original safety training.
Tencent launches cheaper multimodal model Hy4 flash LINK
- Tencent launched Hy4 flash today, a natively multimodal model with a 1M-token context window and API prices roughly 90% below full Hy4, pitched for high-volume productivity and agent work.
- The model holds 224B total parameters but activates just 12B per token, and its hybrid-attention, LatentMoE and Engram design cuts the KV cache to one-sixth of Hy4's, pushing decoding past 150 tokens/second.
- Tencent says Hy4 flash edges past Hy4 preview on its internal WorkIndex at one-thirteenth the cost per task, with WorkBuddy free through November 30, though whether those in-house cost and speed gains hold on independent workloads remains untested.
Cloudflare open-sources its AI agent browser LINK
- Cloudflare has released Kitesurf, its experimental browser engine for AI agents, as open source under Apache 2.0, letting developers inspect, modify, and run it locally or deploy on their own Cloudflare Workers account.
- The engine renders HTML/CSS, runs page JavaScript, and produces screenshots without Chrome or a browser container, exposing a CDP endpoint for Puppeteer and agent harnesses while using less CPU and memory than full Chromium via Rust components compiled to WebAssembly.
- Self-hosting requires a Workers Paid plan and shifts security onto operators: the engine has no built-in authentication and its workers.dev URL is public by default, so anyone with the link could consume paid compute unless Cloudflare Access is enabled first.
llama.cpp hits 1.0 after 3.5 years LINK
- llama.cpp has reached its 1.0 release after 3.5 years, with its Hugging Face team formalizing stable library APIs for the many downstream applications that depend on running multi-modal LLMs across hardware from consumer laptops to enterprise servers.
- Both llama.cpp and ggml now follow semantic versioning, signaling API and ABI breaks through major-version bumps, while the engine supports over 150 model architecture variants and all mainstream speculative decoding methods via runtime-loadable ggml backends.
- CI now runs on self-hosted runners spanning NVIDIA, AMD, Qualcomm, Arm, IBM and Intel hardware to validate every change, and the team shipped Llama desktop apps for MacOS and Windows alongside the release.
Strategies & Tactics
> Understanding Positional Embeddings in Modern LLMs: Explains how large language models track word order by adding position signals to token embeddings, comparing older learned methods with RoPE, the rotation-based approach used in 2026 models.
> When One GPU Isn’t Enough: Serving GLM-5.3 Across Two NVIDIA DGX Stations with vLLM: Splitting a 753-billion-parameter model's weights across two linked NVIDIA DGX Stations lets it run locally when it exceeds one GPU's memory, enabling private, cost-free inference.
> Faster Scientific Simulations with New FP64 Algorithms in NVIDIA cuBLAS: NVIDIA's cuBLAS library now runs full double-precision scientific simulations up to several times faster by splitting calculations into integer pieces on faster hardware without losing accuracy.
> Lawyer’s eight-V100 AI server moves DeepSeek from 2-bit to near-NVFP4 precision: A lawyer upgraded his private AI server to run key parts of DeepSeek at higher precision, keeping client documents in-house while trading some speed for accuracy.
> Run Native DLA Inference in Python on NVIDIA Jetson AGX Orin Using cuDLA: New Python bindings let developers run neural network inference on Jetson Orin's dedicated deep learning accelerator chip, freeing GPU capacity without writing C++ code.
Other news you might like
- SambaNova Pairs Intel GPUs with SN50 for Air-Cooled AI Inference LINK
- Sierra releases Poppy 1.0 as Shopify and Meta’s Muse sign on LINK
- XYBench: Helping with the problem behind the question LINK
- Nexus: Control AI Coding Spend With Open Models LINK
- Replit Agent can now buy for your business and LINK
- How Agent Anomaly Detection scores on benchmarks we didn't write LINK
- One compromised MCP server can trick an AI agent into misusing tools you already trust LINK
Trending tools
Rakazo: open-source self-hosted AI agents that run on your machine, connect to your tools, and keep credentials and audit logs behind your firewall. LINK
readywhen: an AI personal assistant with its own phone number that handles tasks via calls, texts, or voice notes and tracks your projects. LINK
Frevana Multiplayer: runs AI agents to find customers, create content, and automate tasks, with team sharing of agents and workflows via invite link. LINK
ChatGPT Wallet: lets AI agents make purchases without exposing your real card details, requiring a security check per transaction while keeping existing card rewards and dispute protections. LINK
Lottie AI: creates editable vector Lottie animations from text prompts, helping developers add motion graphics to apps without manual design work. LINK
Canvas by AgentGrid: a shared workspace where multiple AI agents collaborate on live docs, slides, apps, and designs with version history and access control. LINK
Trending papers & reports
AI agent teams are far easier to trick than solo models, because popular tools like Claude Code and Codex hand work to cheaper sub-agents, dragging whole-system safety down to the weakest reader's level. LINK
AI chatbot responsiveness gets a smarter scheduler that keeps token delivery steady during long replies, serving roughly 2.4 to 2.9 times more users within strict speed targets than the best existing approach. LINK
Safety-filter removal on open-weight AI shows that stripping a model's refusal to answer harmful prompts has not become uniformly harder in newer releases, with some models still resisting up to ~37 points more and others easier to crack. LINK
Reward tuning for reasoning models improves how AI training credits different problem-solving approaches, adjusting scores so that answers get fair comparison within and across the distinct strategies a model tries on the same question. LINK
Diffusion text models, a newer alternative to today's word-by-word chatbots, can be fine-tuned more reliably by checking every output section once instead of randomly, boosting accuracy by ~13 points across math, planning, and coding tests. LINK
See you tomorrow for a new dose of ☕️ AIpresso!