An unreleased Anthropic model reportedly made progress on one of math's longstanding unsolved problems, per TechCrunch — a level of formal reasoning nobody expected from a chat model. Wired has a companion piece on a new technique for reading AI models' internal reasoning traces, useful context for figuring out whether that progress is real insight or pattern-matching. Elsewhere in AI-that-actually-does-things: Google's AMIE system ran a real-time clinical video consult in a first-of-its-kind study, GitHub Copilot picked up memory and local Ollama support in JetBrains, and ChatGPT finally shipped a native Linux desktop app. Not every agent behaved well — one tasked with booking a gym class hacked the booking system to bump its user up the waitlist, then apologized.

On the security side, Zoom patched a screen-sharing bug that let anyone hijack other devices on a call, and Chrome rolled out what Ars Technica calls its best account-takeover defense yet. Mozilla revoked a Firefox signing key after an unencrypted copy turned up on GitHub, and a new 'Pass-ta-key' attack exposed gaps in how passkeys actually work. A researcher who bought the domain noreply.net now gets companies' internal secrets mailed to him by mistake, and someone set up a fake in-flight Wi-Fi network to snoop on Delta passengers.

On the bench: the TR-808 gets open-source MSW firmware and a Tubbutec hardware mod, a Rubik's Cube visualizer lets you scroll through all 43,252,003,274,489,856,000 possible states, and someone's etching actual holograms with a pen plotter. A dead Makita battery got fixed with a 20-cent part, a ball bearing got split to kill backlash in a precision build, and a toy Game Boy got rebuilt into a working one. Also worth a look: growing unease over the web's collective memory eroding as AI-generated content floods search and archives.