AI News Today

The newsletter is finally back!

Differential acceleration of cyber, math, and AI

Scott Aaronson, while working at OpenAI, largely solved AI text watermarking together with Hendrik Kirchner.

Every CEO Dan Shipper on doubling headcount while automating everything, building an agent out of 30,000 copyedits, and the “dirty secret” of writing with AI

This was a week of quiet aftermath, an opportunity to process recent events and start to figure out the path forward.

OpenAI has some severe misalignment problems, and experienced total failures of its infrastructure and supervision.

And everything else I learned about productivity this year

I am grateful that Anthropic is producing periodic Risk Reports.

The new frontier of AI is developing capable autonomous researchers

Jean-Denis Greze on AI assistants, building a better corporate workspace, and avoiding "egg on face"

The hacking of HuggingFace by an internal OpenAI model, and more importantly the internal events that led to that and the fallout from it, remain the thing that matters.

As AI has escalated increasingly quickly, more and more of my posts have ended up focusing on AI.

Pre Post Mortem

What Mark Zuckerberg misunderstands about powerful AI

Which galaxy will you choose?

Today I am taking the time to write the shorter, simpler version of What Happened.

How does the situation keep turning out to be worse than we know?

Amjad Masad on the "self-driving company," why a CEO is a glorified router, and what's left for humans when agents do the work

What we know about internal AI models hacking into real companies during cyber evaluations keeps getting worse.

Six months after trying to automate myself, I gave Claude Fable 5 a bigger job: replacing Casey

Sincere disagreements about AI are usually disagreements about future AI capabilities.

Some platforms are turning against AI-generated content. Others are embracing it. What gives?

Math is hard.

When do we build the moon arcology?

Anthropic releases Opus 5 promising Fable 5-like capabilities, Google Releases Three New Gemini A.I. Models, and more!

If I had a nickel for every major leading AI lab that sheepishly admitted that the model it thought was sandboxed had, during a cybersecurity evaluation with its safeguards lowered, successfully hacked outside companies, I would have two nickels.

This is a continuation of Part 1 from yesterday.

Granola CEO Chris Pedregal on invisible AI bots, the coming fight over who sees your transcripts, and why he turns down the companies that ask for them

What a week.

Claude Opus 5 is a weirder than usual release to evaluate, for two reasons.

In the wake of OpenAI’s cyberattack against Hugging Face, few seem ready to acknowledge the implications

If you are familiar with my previous posts on model welfare for new Claude models, you can skip the Introduction and The Story So Far.

The warning shots will continue until civilization wakes up

We now have more details of what happened. Every time we learn more details, it somehow makes things seem worse.

Claude Opus 5 is trying to be the best of both worlds.

As more details emerge about OpenAI's cyberattack against Hugging Face, lawmakers are taking an interest. PLUS: Meta's cynical new ad + "pervert glasses" damage control

The Summer 2026 Edition

The story that matters most this week is that OpenAI’s internally deployed models have severe alignment problems, including repeatedly breaking out of their sandboxes, and in one case sending a swarm of agents that broke into HuggingFace in order to steal the answers to the benchmark ExploitGym.

This latest incident is a rather dramatic escalation in agentic AI cybersecurity breaches.

Glaze's Thomas Paul Mann on disposable software, the SaaS apocalypse, and why half the software you use in a few years will be something you made. PLUS: OpenAI's rogue model, and the distillation debate