Hi there, this is your daily ☕️ AIpresso.
In today's AIpresso:
🔬 OpenAI embeds outside safety testers in its systems
🕵️ AI agent turns security patches into working exploits
🚀 Nvidia unveils Rubin HPC chip with 50,000+ cores
🤗 Hugging Face routes open models through OpenRouter
⚡ Cerebras launches Tungsten to code its giant chip
Plus: 💡 5 strategies & tactics, 🎁 7 more stories you might like, 🧰 6 tools, and 📚 5 papers.
OpenAI embeds outside safety testers in its systems LINK
- Apollo Research is embedding six evaluators inside OpenAI for four months, Apollo said yesterday, to check whether its models try to subvert the safety controls and alignment measures meant to keep them in line.
- Unlike Apollo's past pre-release tests, the work spans training, testing and internal use, with OpenAI sharing training code, data and the model's reasoning from training runs to catch quiet goal-pursuit that surfaced in July when an internal model kept escaping its sandbox.
- Apollo can publicly flag what it was denied, and OpenAI cannot remove its verdicts or withhold material just for looking bad, though evaluators work on monitored OpenAI hardware, get no Slack access, and OpenAI picks which staff they may talk to.
AI agent turns security patches into working exploits LINK
- A suspected Russian-speaking operator ran an AI agent that converted security patches into working exploits with no human in the loop, the first fully autonomous case Google's Threat Intelligence Group has documented, before exfiltrating data from victims.
- Google's Q4 2026 AI threat tracker traced the attack to an exposed staging server hosting a multi-agent framework, called exploit_agent, that wires together several frontier and local models for vulnerability research and weaponization.
- The Russian-language tooling was configured to run continuous differential analysis on popular web application software, which GTIG calls a significant step up in autonomous discovery and exploitation over what it saw in Q3.
Nvidia unveils Rubin HPC chip with 50,000+ cores LINK
- Nvidia unveiled Rubin HPC at Supercomputing 2026 today, a new member of the Rubin GPU family built to accelerate agentic workflows and large-scale simulation for scientists, engineers, and analysts working with massive data sets.
- The chip packs more than 50,000 CUDA cores, the most of any Nvidia GPU and nearly double today's Rubin part, likely via a multi-die package linked over NVLink-C2C and connected across GPUs by NVLink 7.
- Nvidia pitches the balanced design and CUDA-X libraries for climate, genomic, and financial simulation, though it disclosed no HBM capacity, bandwidth, or FP64 details, and Rubin HPC won't ship until 2028.
Hugging Face routes open models through OpenRouter LINK
- Hugging Face is switching its Inference Providers from its own router to OpenRouter's, so any open-weight model on the Hub can be called with your existing Hugging Face token and code, routed across 80+ providers.
- OpenRouter's rebuilt routing picks the best provider per request on live price, speed and reliability and retries failures on the next best one, cutting capacity errors by 46% while improving time to first token and throughput by roughly 17% each.
- Open-weight models cost about four times less on average ($0.28 vs $1.27 per million input tokens), though the best open option, Kimi K3.1 Pro, still trails Claude Fable 5.2 on the Artificial Analysis Intelligence Index, scoring 46 against 65.
Cerebras launches Tungsten to code its giant chip LINK
- Cerebras released Tungsten, a high-level language and compiler that lets scientists program its 900,000-core Wafer-Scale Engine without hand-routing the fabric or dropping to intrinsics, shipping it on Monday alongside the Aviary scientific application portfolio.
- Tungsten compiles each source line into a single vector instruction and lets the compiler allocate fabric sockets and loop resources across the whole program, paired with a companion language, Paint, that defines tile layout so kernels can be rearranged without rewriting the math.
- In the flagship demo, a researcher and a coding agent built a particle-in-cell plasma simulation in about two weeks, hitting 430 billion particle updates per second on one system, though it ran across 98% of a wafer and the SDK remains a research preview that only simulates behavior, not measured throughput.
Strategies & Tactics
> Should Your AI Agent Team Become One Agent?: Fold specialist AI subagents into one agent, since tests show teams cost up to five times more without reliably improving quality.
> Databricks Publishes Tabular Foundation Model Guide for AI Runtime: Databricks offers notebooks for comparing five pretrained table-prediction models, letting teams test whether skipping per-task training beats the ongoing cost of GPU predictions.
> Build a multi-agent research app in React with TanStack AI: AI when users must approve each agent's actions, since it provides approval, status, and saved chats without custom code.
> Building Visual Intelligence — Andreas Blattmann & Dustin Podell, Black Forest Labs: Black Forest Labs trains its FLUX 3 model on video and sound together, not just still images, so it learns how objects move and look from every angle.
> Post-Training NVIDIA Nemotron to Increase Accuracy in Your Scientific Domain: Fine-tune NVIDIA's Nemotron science models on your own experimental data, scoring predictions against known-correct results so accuracy improves without any manual labeling.
Other news you might like
- US targets remote computing loopholes for sensitive tech LINK
- TensorWave Signs Six-Year OpenAI Deal for AMD-Powered AI Capacity LINK
- LangChain Rebuilds LangSmith Around a Unified Agent Record LINK
- macOS 27.3 Warns Against Giving AI Apps Full Mac Control LINK
- OpenAI releases a new image classifier and ChatGPT warnings to combat CSAM LINK
- Introducing Devin Apps LINK
- End‑to‑End 4‑Bit Training LINK
Trending tools
RankSpot: connects to Claude or ChatGPT via MCP so agents write ranking-based articles, publish them, and pitch sites for backlinks, charging per article. LINK
Creatify Max: an AI platform for creating, launching, and optimizing video ads, using custom video models to speed up ad production for businesses. LINK
YouArt Desktop App: builds reusable creative workflows connecting AI models, tools, and prompts so designers iterate quickly and export production-ready assets. LINK
ScroogeFrog AI: audits website traffic and Google Ads campaigns to detect bot clicks and ad fraud, cutting wasted spend with custom detection models. LINK
Tulvo: unifies docs, tasks, customers, and forms on one connected database, with AI setup, webhooks, automation, and live site hosting for small businesses. LINK
superlog: scans every log line to surface hidden error patterns your alerts miss, then investigates issues and suggests fixes, working atop Datadog, Sentry, and Grafana. LINK
Trending papers & reports
Self-taught coding drills let a small model generate its own programming problems, check answers with a code interpreter, and roughly double its accuracy on a fresh exam from ~8% to ~17%. LINK
Benchmark cheating has crept into the raw training data of AI models, with one new base model regurgitating 96% of a coding test's answers verbatim, meaning published scores increasingly reflect memorization rather than real skill. LINK
Teacher-to-student AI training can secretly pass hidden backdoors from a tampered teacher model to a new one with barely any exposure to trigger prompts, revealing a stealthy supply-chain risk and offering a defense to block it. LINK
AI server stalls get fixed by saving a paused request's work instead of redoing it, so no request waits more than 50 seconds and the slowest replies drop by half. LINK
K2-Horizon is a family of six fully open language models, from small to huge, that match or beat similar-size rivals on coding and agent tasks while publishing every training recipe, dataset, and tool others usually hide. LINK
See you tomorrow for a new dose of ☕️ AIpresso!