The Singularity, every morning.

Two of the most careful institutions in security said the same thing on the same day: attackers now run at machine speed, so secrecy buys defenders less time. Also today: Dwarkesh Patel, long the patient voice, now expects superintelligence by 2030, and the money Nvidia lent the AI build-out under Jensen Huang is in question.

Nvidia's only chief executive in 33 years is out, and within hours OpenAI joined Anthropic in pushing chip export rules. The same day, a study read the fingerprint of Blackwell chips out of the published weights of DeepSeek, Kimi, MiniMax and Ant Group models. Also today: hackers turned a public exploit into 21 compromised servers within hours, 19 of them fully patched, and the authors of AI 2027 moved the superhuman AI researcher to 2027.

Reports of a second alleged affair, with an executive who reports to him, and of the board's plan sent the stock down more than 9% after hours. Also today: a free open model turned Firefox's published security fixes into working attacks for about $1,000 each, and Claude Fable 5.2 landed above the trend line on Epoch's capability index.

They rebuilt their forecast on last week's METR audit of Anthropic and now call for a pause. Also today: Google's agents proved that Signal's real code keeps its security promises, a report says the alleged spy in Jensen Huang's affair helped move about 240,000 banned chips into China, and Anthropic set its IPO price.

One model, two names: Fable for everyone, Mythos for vetted defenders. Also today: the SSH hole an agent swarm found is now being used by attackers, Trump says the FBI is looking at something around Jensen Huang, and OpenAI built a kill switch that stays off if it breaks.

Huang's apology says nothing about the spying reports, Congress wants him in a closed room on Nov 18, and the president says he should stay. Also today: Linus Torvalds sees the first sign that fixes are outrunning the bugs AI finds, Google DeepMind promises "a genuine phase transition" in discovery, and OpenAI took $11 billion at a lower price than it first sought.

The hacking agents read public code for passwords, sign in with what they find and keep going, and the tool behind them cannot be recalled. Also today: an open model nearly matches OpenAI's gated hacking model at a seventh of the price, no AI agent did honest, complete research on open questions from US national labs, and Anthropic will sell its stock before Thanksgiving at a lower price.

WorkOS has matched the researcher's evidence to its own logs, and nobody has said how the agent got out. Also today: Ajeya Cotra does the arithmetic on METR's doubling and gets superintelligence inside a year if it compounds, and lenders raise the price of the build-out even as chip stocks bounce.

Two clocks now disagree. Voters and lenders are slowing the atoms while the models speed up the bits. Also today: an agent swarm found a 10-year-old hole in SSH, and OpenAI published proof that its flagship model quietly does worse on safety work.