Two of yesterday's stories ended up in the same tab group for me. In one, a mathematician says he proved the Riemann hypothesis with Claude, and strangers on a Zulip server can check it by compiling a file. In the other, a security researcher sent nearly two million prompts into the AI inference market and watched his secrets come back out the other end.
These two stories mark the line I keep running into. AI output is getting easier to verify. The plumbing it runs through is getting harder to see into. Everything else from yesterday lands on one side of that line or the other.
The proof you can compile#
Levent Alpöge, a mathematician who works at Anthropic, published an essay claiming a proof of the Riemann hypothesis, the 1859 problem about how prime numbers are distributed. It is one of the seven Millennium Prize problems. He says he found it working with Claude, and that the method extends to the Generalized Riemann Hypothesis. TechCrunch reports that the proof is written in Lean. Several people in a Zulip chat, including MIT mathematician Andrew Sutherland, say they compiled it successfully.
That detail matters more to me than the headline. For most of history, a claimed proof of something this big meant months or years of a few experts squinting at a manuscript. Now the first question isn't "do we believe him?" It's "does it type-check?" Anyone with a laptop can get that answer in an afternoon. Understanding the argument is a separate and much slower job, and Alpöge admits it. He says he felt guilty releasing a proof without a fully human-readable explanation, and has committed to writing one.
Keep the caveats in view. The Verge notes that Anthropic hasn't commented, and that the Clay Mathematics Institute requires two years of scrutiny and "general acceptance" before it pays out the $1 million. Back in August, Anthropic itself said it didn't expect Claude's techniques to lead to a Riemann proof. On the same day, Stanford postdoc Alessandro Masullo posted a proof of the Generalized Riemann Hypothesis. Ravi Vakil, the Stanford professor who posted it, described it as "all human, with no AI." It took 12 years, and it isn't in Lean yet.
A formal proof that compiles is the strongest kind of AI output there is. You don't have to trust the model, the author or the company. You trust the checker.
The pipeline you can't#
Now the other side. Ars Technica has the details of HD Moore's experiment. Moore is chief research officer at Dragos. He seeded unique, never-used secrets into prompts and sent them through AI routers and resellers: working AWS keys tied to empty accounts, SSH passwords, private URLs and git repository references. Each secret went into exactly one prompt to exactly one provider, so any leak could be traced to a single hop.
~2 million prompts, each carrying a trackable canary
3 hosting providers found breached. In one breach, every request sent during the affected windows was stolen.
Tens of thousands of AWS API calls made with the stolen bait keys. Attackers probed S3, enabled Bedrock models and created their own IAM users.
More than half of prompts sent to one Chinese lab's models had their secret URLs fetched later. That included about two-thirds of the requests flagged as zero data retention (ZDR).
A privacy flag doesn't help when the attacker is already inside. — HD Moore, via Ars Technica
The detail that should worry builders most: the same scraper showed up when prompts went through US hosts that resell that lab's models. Picking a US provider to keep your data in the country didn't keep it there. Some resellers offered low prices, switched on request logging and simply waited.
Here is what I'm changing this week. No credentials in prompts, ever, including the "just debugging" ones that end up in agent context. No accepting a router's default provider pool. You pick the upstream yourself, and you treat the ZDR flag as a request, not a guarantee. Moore's own ranking is self-hosting first, then a direct contract with one first-tier provider, then everything else. That's the right order.
Signal and METR hit the same wall from opposite sides#
Signal is blocking screenshots by default on iOS 27 and Android, so Siri, Gemini and other OS-level assistants can't read your chats off the screen. Meredith Whittaker is blunt that the fix is "incomplete." Agents can still get at app data through the file system, and Signal can't ship the protection on macOS at all, because Apple, Google and Microsoft don't give developers proper controls or documentation. Signal's complaint is the same as Moore's, one layer up: it can't see what the agent is doing, so it can't promise anything to its users.
METR's Q4 capability update, covered by MIT Technology Review, arrives at verification from the evaluation side. Coding tasks, which are easy to check, keep climbing on the exponential curve. The new judgment-heavy tasks don't. Asked to design evaluations for METR's own research, the best model (Opus 5.5) only reached "borderline above the bar." None of the outputs was useful enough for METR staff to actually use. On scoring philosophical arguments, models fell short of humans. METR's Matt Putz put it plainly: hard-to-verify tasks are "much harder to study, and especially to quantify."
Read next to the Riemann story, that's the whole picture. Models do best exactly where a machine can grade the answer. Math with Lean and code with tests are where progress looks unstoppable. Where the only judge is a person's taste, the curve flattens out, or at least we can't measure it well enough to say.
Google removes a button nobody pressed#
A smaller story, but it points the same way. Google's desktop homepage no longer has a Google Search button. In its place are buttons for image generation, file analysis and brainstorming with Gemini, plus a sidebar of your recent AI Mode threads (The Verge also covered it). Practically, almost nothing is lost, since everyone hits Enter. As a signal, it's clear: the default entry point to the web is now a conversation that Google hosts and remembers.
Trust what you can check#
My prediction: within a year, "verifiable" becomes a product feature people pay for, at both ends of the stack. On the output side, that means more Lean, more test-gated agents and more evals that grade themselves. On the infrastructure side, it means attested inference, signed provider chains, and OS permission models that actually tell an app which agent read what. The Riemann proof shows how good it feels when you don't need to trust anyone. Moore's canaries show what happens when you do.