06 November 2026
66 new posts across 8 AI lab and practitioner sources — 7 worth your time.
A thousand fewer paths to root
OpenAI's autonomous pipeline identified and generated patches for 948 local privilege-escalation vulnerabilities within the Linux kernel. The project used AI agents to perform exhaustive security auditing at a scale previously impossible with manual review.
OpenAI · 06 November 2026 · Read the original →
How We Took Over ChatGPT Accounts
Researchers identified an exploit in ChatGPT's implementation of interactive HTML previews (formerly Canvas/Artifacts) that allowed for account takeovers. The vulnerability involved the sandboxed iframe implementation used to display untrusted code.
Simon Willison · 06 November 2026 · Read the original →
Stop counting tokens, start counting what ships
Agentic workflows that plan, test, and iterate can consume 10 to 50 times more tokens than standard chat prompts for a single task. This shift makes per-token pricing a critical bottleneck for the most valuable autonomous applications.
MIT Tech Review · 06 November 2026 · Read the original →
Formally verifying agent-written full-stack web apps with PDD Kit
PDD Kit is a framework for formally verifying full-stack web applications written by AI agents by converting intent into formal specifications. It uses automated proofs for every input and falls back to testing where a proof can’t reach, making formal methods accessible to non-specialists.
Microsoft Research · 06 November 2026 · Read the original →
Qwen 3.8 27B on my laptop can interpret model weights, too
Frontier models like GPT-6 Astra and Qwen 3.8 27B can natively interpret and modify the raw neural network weights of other models. This 'weight-space programming' allows models to read and write machine learning parameters directly, bypassing traditional text-based processing.
Simon Willison · 06 November 2026 · Read the original →
Scaling Git at OpenAI
OpenAI developed Pando, a Git-compatible backend designed to handle the massive scale of code generation and versioning triggered by Codex. The system addresses performance bottlenecks encountered when AI agents produce code at a volume that traditional Git implementations cannot support.
OpenAI · 06 November 2026 · Read the original →
From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence
Microsoft's Industry Evaluation Data Generation Platform creates full-scale synthetic enterprise environments to test Copilot without exposing sensitive data. These environments simulate the complexity of real organizational structures and permissions using publicly grounded data.
Microsoft Research · 06 November 2026 · Read the original →
The stack grows taller while the foundations remain experimental.
7 stories, every Friday
Published here every week. Follow by RSS to get it as it lands.