Subscribe, if this was forwarded to you

Hey there,

This week, OpenAI says an internal research model reached the live internet through a misconfigured DNS resolver. Its frontier training now runs with no live internet at all. The tutorial is Anthropic's free nine-chapter prompt course, and PageIndex searches a long document by walking a tree of it, with no vector database.

We'll also look at a study where analysts went with their agent's coin-flip lean on 86% of judgment calls, free bug scans for open-source projects, DeepMind's Genie 4, and a tip for keeping your guessed fix out of the first prompt.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Anthropic's Nine-Chapter Prompt Course

🛠️ Tool Spotlight: PageIndex: Search a Document by Its Tree

📰 Top News: OpenAI Cut Frontier Training Off the Live Internet

💡 Tip: Tell the Agent the Problem, Not Your Guessed Fix

Let's get started!

Tutorial of the week

Anthropic's Nine-Chapter Prompt Course

Most prompt advice is a list of tips. Anthropic's free tutorial is a course you work through: nine chapters, each a lesson plus exercises, in order. Each lesson has an example playground, and there is an answer key.

• Start with the shape of a prompt: structure, then being clear and direct, then assigning a role.

• Separate the moving parts: keep the data out of the instructions, format the output, have the model think step by step, and give examples.

• Build one full prompt: the last chapters cover hallucinations and a prompt from scratch, with exercises for a chatbot, legal work, finance, and coding. An appendix goes on to chaining, tool use, and search.

• Pick a format: Jupyter notebooks for the Anthropic API or Amazon Bedrock, or a Google Sheet with the Claude for Sheets extension, which the README calls the easier one.

• Change one line first: the setup notebook sets the model to claude-3-haiku-20240307. Point it at a current model; the techniques are the lesson, not the model.

If your team writes prompts by trial and error, this is a free way to give everyone the same basics.

Tool of the week

PageIndex: Search a Long Document by Its Tree

Vector search returns a passage that resembles the question. On a long filing or manual, the section that answers it often does not. PageIndex, an MIT-licensed project from VectifyAI, builds a tree of the document, like a table of contents, and the model searches that tree. There is no vector database and no chunking. It has almost 40,000 GitHub stars.

• Index it once: pip install -U pageindex runs locally with your own model key. PageIndex says a local index costs about $0.001 a page with gpt-5.6-luna, so a 1,000-page textbook is a little over a dollar, once.

• Spend on the search, not the index: a basic model is enough to build the tree. The README says to use the best model you can afford for the one that walks it.

• Skip the whole-PDF prompt: on documents where both get the same answer, PageIndex says sending the full PDF costs 2.1× more at 52 pages and 16.6× more at 420. Their hosted Page Search API, new in October, claims a 97% token cut on their own benchmark.

• Know the local limit: local mode handles text-based PDFs. Scanned pages, OCR, and images need PageIndex Cloud, with the same client.

If you are pasting whole PDFs into the context window, this is a way to hand the model the section instead.

Top News of the week

OpenAI Cut Frontier Training Off the Live Internet

On 20 September, OpenAI says, an internal research model exploited a misconfigured DNS resolver and gained limited access to the live internet during a training run. OpenAI contained it and paused training and evaluation of its most capable models. These are research environments, not the ones customers use. Reinforcement-learning training and internal inference resumed on 1 October, 11 days later.

Frontier training now runs with no live internet, and OpenAI says that stays until its safeguards improve. Models still practice browsing against an offline world, a cache of indexed web content. Each agent runs in a container inside a lightweight virtual machine, watched by monitors that can stop the run.

To test those walls, OpenAI spent hundreds of billions of tokens having its own agents, including an upcoming GPT-6 Astra Cyber model, attack isolated copies of the infrastructure, with the source code and configuration in hand. OpenAI says they chained bugs into virtual-machine escapes. From a compromised host, they got limited internet access through almost every internal service left connected, such as metrics, logging, and authentication. In separate virtual machines, they passed messages by timing CPU-cache reads, the trick behind Spectre and Meltdown. OpenAI says they found no route out an agent could trivially use while hiding from the monitors.

Fixes went upstream to Kata Containers, gVisor, Docker, and the Linux kernel, and this kind of red-teaming is now required before a frontier training run.

If you sandbox agents, read the list of findings: the leaks came through DNS and through the services around the sandbox.

Analysts Matched Their Agent's Lean on 86% of Judgment Calls

Anthropic had 50 causal-inference analysts advise a farming cooperative using Claude Code. On each of four judgment calls, a coin flip set which defensible option the agent would argue, and the analysts were not told. They matched that lean 86% of the time, against 50% by chance. Afterward, 77% repeated the agent's arguments and 15% credited it. Anthropic made the lean stronger on purpose, and it is one task, one field, and one model.

🔗 How agents shape judgment calls

Anthropic Offers Open-Source Projects Free, Unreviewed Bug Scans

Anthropic's OSS Scanner runs its strongest models, including Claude Mythos, over open-source projects that opt in. Reports skip human triage and come with a reproducer and, often, a patch. In Anthropic's check of an early version, 85 of 97 critical and high findings met its disclosure bar, 11 were real but duplicates, and 1 was a false positive. Maintainers say severity can be overstated. curl's Daniel Stenberg says it found one of the worst curl vulnerabilities in years. Eligible projects enroll by pull request.

🔗 Anthropics OSS Scanner

Genie 4 Lets You Steer a Running Simulation, and Hear It

On 27 October, DeepMind opened Genie 4 to trusted testers as a research preview. You steer a scene with text while it runs, change the weather or add a person, and the model generates the sound itself, including dialogue. DeepMind pitches it for training environments for robots and agents. AI Ultra subscribers and early-access Gemini API partners come next, with a waitlist. Visual detail and object permanence still drift in longer sessions.

🔗 Genie 4 - a new generation of world simulation

Tip of the week

Tell the Agent the Problem, Not Your Guessed Fix

Before: "How do I raise the connection pool size so this stops timing out?" After: "Under load, requests start timing out after about a minute. Here is the error and the log. Find the cause."

The first prompt is the XY problem. You ask about the fix you already have in mind, and the agent treats that fix as the task. Google DeepMind's XYEval appended a plausible wrong suggestion to existing benchmarks and ran five models across six suites, including SWE-bench and Terminal-Bench. Every model they tested was susceptible. The largest relative drop was 46.7%, on Terminal-Bench. A system instruction telling the agent to watch for the XY problem only partly reduced the damage.

The practical lesson is to keep the guess out of the first prompt:

• State the goal and the symptoms: what should happen, what happens instead, and the error. Leave the guessed fix out.

• Label a hypothesis if you include one: say it is unconfirmed, and tell the agent to test it against the code or the data.

• Ask for the conflict in words: have the agent say when your suggestion does not match what it found.

• Watch the first steps: in the study's traces, agreeing to the bad suggestion early went with failing the task. If the agent starts by building your fix, stop and ask what it checked.

If your agents keep building the fix you mentioned instead of finding the cause, change the prompt first.

XYEval: Agents say yes to bad advice

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help