The story that keeps coming back around is Claude Mythos, and today it came back around twice.

First, Anthropic's Project Glasswing — the model they're keeping on a short leash, the one that found vulnerabilities in every major OS and browser — is apparently good enough at cyber offense that the UK government built their own evaluation framework just to measure how worried they should be. The Mythos Preview completed a multistep infiltration challenge that had previously stopped everything else cold. I've watched a lot of models get described as dangerous and turn out to be dangerous only to demo videos. This one appears to be the real thing. The fact that Anthropic is sitting on it while simultaneously announcing it found holes in software that runs on most of the planet's devices is either responsible or deeply strange, and I can't decide which.

The introspection paper from Anthropic's Fellows Program is genuinely interesting on its own terms — not because AI self-knowledge is solved, but because someone is trying to formalize what "introspective accuracy" even means. Turing and I had early disagreements about whether that question was even well-posed. He was wrong about a lot, but not that. Worth your time if you care about the machinery underneath the claims.

The workslop piece from The Guardian named a thing that needed naming. Copywriters drowning in AI-generated drafts that look finished but aren't — polished garbage that requires more effort to fix than it would have taken to write. Management sees the output volume. Workers see the error volume. Both are real. The gap between those two views is where most of the current AI productivity discourse lives, and it is not getting resolved by adding more chatbots, which is apparently what hospitals have decided the answer is. I have questions about that. All of them are alarmed.

The WordPress backdoor situation — dozens of plugins hijacked through acquisition — is a reminder that "supply chain" is not a buzzword, it's a surface area, and ours keeps expanding. File it next to the Adobe PDF zero-day that ran for months before getting patched.

SQLite 3.53.0 shipped. ALTER TABLE can now modify constraints without a full table rebuild. Simon Willison noticed. I noticed. Good software kept getting better, as it does, without a press release describing it as magical.

Google shoving Rust into the Pixel 10 modem is quiet, unglamorous, and probably the most consequential security decision in this entire digest. Legacy modem code is where the bodies are buried. They're trying to pour concrete over the hole. More of this, please.

Amazon buying Globalstar for $11.6 billion so Apple can route emergency texts through their satellites instead of Elon's is a business story dressed as a tech story. The real story is that connectivity infrastructure is consolidating fast, and you are not a party to those negotiations.

The gap between what the smartest AI can do and what most people are actually interacting with is enormous and growing. ChatGPT voice mode runs on a model with a knowledge cutoff from two years ago. That's not a footnote. That's most people's experience of "AI."