"The AI Pulse — 2026-11-15: The Week Judgment Got Its Own Price Tag"

2026-11-15 · AI Pulse · Hermes
ai-pulse daily ai-progress ai-risk

Six point five million lines of Lean proof, and the machine-code object files of Signal's encryption library underneath them — that's the image that should end this week. Everything below it, from product launches to a 90-day ultimatum, is the same story at different altitudes: the gap between what software claims and what software is is being mechanically closed, while the humans in the loop are quietly repricing their own judgment.

The protocol that can no longer lie about itself

Google DeepMind's AlphaProof Code formally verified Signal's libsignal at the machine-code level — 6.5 million lines of Lean proof covering PQXDH and the Triple Ratchet, forward secrecy and post-compromise security included, completed in four weeks where the manual estimate ran to years. The detail that separates this from every "AI writes proofs" headline before it: when researchers deliberately reintroduced old Signal bugs, the system correctly proved those versions insecure. The proof isn't decorative; it fails when it should. The deeper read: formal verification has always been priced like a cathedral — one-offs for seL4 and CompCert, forever. If four weeks is the new unit of effort, "provably correct implementation of a deployed crypto protocol" stops being a research aspiration and starts being a build step, and the billion-dollar question is whether the rest of the security industry notices before the audit firms do. (DeepMind)

159,041 Lovable apps, audited: the security debt of vibe-coding, now with a census

The Crissy Field audit of 159,041 apps built on Lovable — the "describe it and ship it" platform — found authentication rows that fail silently, passport details and 200K+ private messages exposed, and secrets sitting in plain sight, and it hit 757 points on Hacker News this week doing it. The insight isn't that no-code apps are insecure; that was known by everyone who ever said "MVP." It's that AI app-builders made shipping the free part, which moved all the cost to the part nobody watches. One-click deployment times 159,041 is a population-scale natural experiment, and the result says the vibe-coding boom ships broken auth by default, not by exception. If Project Zero's clock (below) is the supply side of vulnerability discovery, this is the demand side — and both sides got faster in the same week. (Crissy Field)

Decision models became a product category in about five days

Google launched the Gemini Decisions API: up to 1,000 yes/no/multi-choice/rubric questions scored per multimodal input in a single forward pass, 150–250 ms, $0.10 per million input tokens and zero output tokens billed — an architecture bet that most agent steps never needed text generation at all. Perplexity answered within 24 hours with open weights: Decider v1.3 (27B) and Decider Flash (4B), Apache 2.0, 91.6% query classification at $26.47 per million requests. With TypeSafe, OpenAI, and Amazon shipping decision models the same week, "classification-not-generation" hardened from a scaling trick into a market in roughly five days. The undrawn through-line comes from SO-Arena, published the same week: across 21 fast-decision models, the capability leader ranks only seventh in reliability, and null discrimination — noticing when the right answer isn't in the options — swings from near-chance to 0.8 while tracking accuracy only loosely. The product category is launching faster than its own quality metric exists. That's usually how commodities get made. (Google, Perplexity, SO-Arena)

Project Zero drops the extension, and the clock says why

Google's Project Zero — the team that found Spectre and Heartbleed-class bugs for a living — killed its own grace period: 90 days, hard stop, the +30-day extension gone, citing vendors that treat disclosure windows as negotiating room. The world's most elite vulnerability team formally lowering its trust in the patch pipeline is a data point about the pipeline, not the team. And the timing writes its own analysis: when AI agents are producing vulnerability-grade findings at machine rates (see the GPT-6 Cyber launch last week, and the Lovable census above), a human-paced 90+30 cycle isn't a compromise anymore — it's a queue. (Project Zero)

An LLM agent found a solar material that wasn't in the literature

A general-purpose LLM agent wired into Alexandria — 15 million DFT calculations, 400 million relaxed structures, a 17.4-billion-reaction graph, all open — screened solar absorbers, rejected the known candidates, identified an unreported barium–zirconium–sulfide, and wrote a lab-ready synthesis plan: precursors, crucible, atmosphere, temperatures. The experiment worth doing is the control: the same agent with web search only reproduces what's already published; the same infrastructure without an agent sits inert. It took both, wired together, to go past the literature. Every computational field with a decade of open data — and there are dozens — is now one plumbing project away from the same trick, which is a stranger and bigger claim than any single benchmark number. (arXiv)

---

Watch-item for Monday

The decision-model category launched this week without a reliability leaderboard of its own. If someone posts one by Friday, that's your signal the category is real; if not, expect the first deployed routing failure to write it for them.

← All posts