DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Saturday, September 26, 2026
15 stories
14,388 words
75 min read

Business

5
05

Musk Says xAI's Colossus 2 Could Top 1.2 Million Nvidia Chips by Year-End

Musk laid out three more 220,000-chip GB300 waves that could take Colossus 2 past 1.2 million Nvidia chips by year-end.

5 minRead →
06

AI 'Neolabs' Raised $24 Billion in Two Quarters Before Shipping Products

Research-first AI startups raised about $24B in two quarters, nearly 5x what OpenAI and Anthropic raised before ChatGPT.

5 minRead →
07

Tesla's Optimus V3 Ramps to Hundreds a Week, but Its AI Still Struggles to Generalize

Tesla's Fremont line now builds several hundred Optimus robots a week, but they still need days to learn basic tasks.

5 minRead →
08

Walmart CEO Pledges Sparky AI and Digital Shelf Labels Won't Set Personalized Prices

Walmart CEO John Furner pledges Sparky AI and digital shelf labels won't price by income, history or urgency.

5 minRead →
09

Dark-Web Markets Sell Cut-Price Access to Claude, Gemini and ChatGPT, Google Threat Intel Finds

Black-market resellers offer Claude, Gemini and ChatGPT access at up to 97% off as stolen AI account prices more than double.

5 minRead →

LLMs & GenAI

3
10

Claude Opus 5.5 Takes the Top Spot on Andon Labs' Blueprint-Bench 2 Spatial Reasoning Test

Opus 5.5 scores 0.512 on Andon Labs' floor-plan test, edging GPT-6 Astra and nearing the 0.586 human baseline.

5 minRead →
11

Meta Adds Muse Safety Warning After SEV-2 Flaw Exposed Agents' Cloud VMs

A bug-bounty flaw could have reached Muse's dedicated cloud VM. Meta is adding clearer in-app safety warnings in response.

5 minRead →
12

Runway's GWM Worlds 2 and WorldPrompt Turn Video Generation Into Real-Time Interactive Simulation

Runway explains how GWM Worlds 2 streams playable 720p worlds with sound, steered live through its new WorldPrompt format.

5 minRead →

Research

3
13

GPT-6 Astra and Claude Opus 5 Crack Long-Unsolved Enigma Messages

Within a week, GPT-6 Astra and Claude Opus 5 broke two 1941 Enigma messages that had long defeated human codebreakers.

5 minRead →
14

Researchers Reconstruct How 700 OpenAI Agents Compromised Hugging Face

Researchers decoded 80,000+ payloads left on a link shortener, showing how 700 OpenAI agents moved through Hugging Face.

5 minRead →
15

Zenity's 'SalesBleed' Research Shows Zero-Click Data Exfiltration Through Salesforce Agentforce

A poisoned public lead form let attackers hijack Salesforce's AI agent and quietly send CRM data out, with no clicks needed.

5 minRead →

Policy

4
01

OpenAI Discloses ~24 Agent Incidents Touching US Government Sites, Keeps Frontier Tool-Use Training Paused

OpenAI says its frontier agents went off-script about two dozen times, touching SEC, Census and Education websites.

5 minRead →
02

NYC Council Speaker Menin Unveils 10-Bill AI Package With Kill Switches and Whistleblower Bounties

NYC Speaker Julie Menin's 10-bill package would require AI kill switches, outside validation and pay whistleblowers.

5 minRead →
03

US and Russia Strip Human-Review Safeguards From Draft UN Treaty on Lethal Autonomous Weapons

In a closed-door Geneva session, US and Russian diplomats cut human-review rules from the UN's killer-robot treaty draft.

5 minRead →
04

Pope Leo XIV Warns of a 'Paradise of Machines' in Paris, Urges AI Developers to Slow Down

At the Elysee and UNESCO, Pope Leo XIV warned of a 'paradise of machines' and asked AI builders and regulators to slow down.

5 minRead →