The Encrypted CoT Block Had One Global Key. Every Session Shared It.
Researchers turned one global encryption key into a cross-user chain-of-thought extractor, recovering 182 credentials from 315,000 publicly logged reasoning blocks.
Category
Editorial
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
96 results in this archive
The Known Proportion of Riemann Zeros Just Jumped 25 Points
Claude improved the known lower bound for zeros of the Riemann zeta function on the critical line from 41.6% to 67.2%. Conrey and Goldston read the proof. Here is what the result actually says.
You Can’t Reason Your Way to a Selection Pressure
I gave the plants in a small world a genome, and for most of a day they evolved in the wrong direction. Three times I designed the selection pressure, and three times the pressure I specified and the pressure I built were different objects. Only the measurement showed the gap.
Someone Held My Site for Twelve Days. They Were After the Kernel.
Four webshells and six administrator accounts I did not create, from a WordPress core flaw I had published an article about five days earlier. The patch existed and was pushed to the whole internet. It never reached me, because four months before I had silenced the scheduler that installs patches while trying to make the site safer.
OpenJDK Banned AI Contributions. GraalVM Didn’t. Both Are Oracle.
Oracle controls two major Java projects with opposite AI contribution policies. The difference isn't quality — it's liability.
Suno Scraped YouTube. Germany Found the Music Inside.
Europe's first AI music ruling found what Suno denied: the training data memorized the songs. Now every platform knows what legal exposure looks like.
AMD Burned a Model Into Silicon. The Bet Is That Models Stay Put.
Taalas's HC1 eliminates inference's memory wall by etching weights into transistors. AMD's acquisition is a bet that production inference stabilizes — and the cost math makes it defensible.
The Builders of Google AI Left to Automate What They Built It For
Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals just left Google to automate the research loop itself. That is a directional signal, not just a talent event.
The Guard Model That Takes Its Policy at Runtime
Mistral's Shieldstral accepts its moderation rules as a plain-text runtime argument. Every other major guard model has those rules in its weights. That's a meaningful architectural shift — and the benchmarks don't capture what it actually changes.