DeepMind Proved Signal's Code Is Secure — Demand the Same Standard From Your AI Vendors
On November 10, 2026, Google DeepMind published a result that quietly resets the standard of evidence for every software vendor you buy from. AI agents generated over six million lines of machine-checked proof demonstra
On November 10, 2026, Google DeepMind published a result that quietly resets the standard of evidence for every software vendor you buy from.
AI agents generated over six million lines of machine-checked proof demonstrating that the actual deployed code behind the Signal protocol — the encryption protecting billions of conversations in apps like Signal and Google Messages — delivers the security properties it claims. Not the design on paper. The compiled code running on devices. Human experts specified what to prove and reviewed every specification; agents powered by Gemini Carbon did the proving; the Lean proof language checked every line mechanically. Four weeks, for a task DeepMind says historically took expert teams years — and which, by their account, no widely deployed protocol implementation had ever been formally verified before, by any method.
Notice what's missing from that claim: "we tested it a lot."
No badge. No attestation. No "trusted by thousands of teams." A proof is different in kind, not degree — it's an artifact you can check, not an assurance you're asked to believe. DeepMind even open-sourced the verification artifacts so others can.
Your AI vendors will not hand you Lean proofs. But the frontier of evidence just moved, and there is a business-owner-sized version of it you can start demanding tomorrow.
Assurance vs. artifact
Almost every AI buying decision a small business makes today rests on assurance: a demo, a security page, a logo wall, a compliance badge. Assurance is a vendor telling you something is true. An artifact is something you can inspect, re-run, or audit yourself — a checkable claim.
The difference matters most exactly where small businesses spend money: tools and agents that touch customer data, send outbound communication, or act on your behalf in systems you don't watch all day. When the tool is wrong quietly, assurance is what you had. When you need to find out what happened, an artifact is what you need.
You don't need a research lab to close that gap. You need three things from the vendor, in writing, before you sign.
The three artifacts to demand from every AI vendor
1. A written scope — what it can and cannot touch
DeepMind's proof started with humans vetting a specification: exactly what was being claimed, and nothing more. That's not a formality — a proof is only as strong as its stated scope, and that's the part worth stealing. Demand one page from the vendor that names the data categories the tool reads, the systems it can touch, and the actions it can take — plus, just as important, what it explicitly cannot.
"The agent only works inside your CRM sandbox and cannot send emails" is a scope. "Enterprise-grade security" is not. If the vendor can't produce the page, the tool's boundaries are undocumented — which means unknown, to both of you.
2. A planted-failure test you can re-run
How did DeepMind convince themselves the system could tell secure code from buggy code? They reintroduced two real, previously fixed bugs into an isolated copy of the Signal code — and in both cases, the system produced a machine-checked proof that the buggy version was insecure. Known answers, deliberately broken.
That's the shape of question to put to any vendor: "Show me a failure your product catches." A support agent should catch a ticket linked to the wrong customer account. An invoice tool should catch a refund for the wrong amount. If the vendor can't show you a single failure their system detects, they haven't instrumented for failure — which means you'll be the one discovering it, in production, with your customers.
A claim you can't test is marketing. This question turns marketing back into a claim.
3. An action log you can read after the fact
The DeepMind agents flagged roughly one percent of the proof for a second look by a human; a machine checked the rest. Volume handled by automation, judgment reserved for people. That division of labor is the right model for any AI tool acting inside your business — but it only works if the tool records what it did.
Demand a log of every action the tool takes on your behalf, in a form you can read without the vendor's help. A log converts trust into verification: when something looks wrong next month, you can see what actually happened instead of filing a support ticket and hoping. If the vendor's answer is "we can pull that for you," that's not a log — that's still assurance, with a delay.
Honest boundaries
- Scope is the proof's strength, not its weakness. The Signal result covers the core protocol library; reporting on the work (The Next Web) notes the guarantees cover one-to-one chats, not the entire Signal apps. An artifact covers exactly what it says it covers — so the closing question in any vendor conversation is always "what does this cover?" Never let a certificate be waved at a question it doesn't answer.
- You won't get formal verification at SMB prices. The standard being imported is the class of evidence — checkable artifacts — not Lean machinery. Scope, test, and log are the small-business translation, and they're achievable now.
- Demanding all three on day one will slow some purchases. Fine. Prioritize: anything that touches money, customer data, or outbound communication gets all three, no exceptions. Everything else, the written scope alone is a reasonable start. A vendor who resists even the scope is telling you something.
The step that comes before all of this
You can't demand artifacts from vendors you don't know you have. Most small businesses are in exactly that position: AI features embedded in software you already pay for, subscriptions nobody opens, and tools the team adopted that never went through anyone's approval. Evaluating evidence requires knowing who your vendors actually are — start with a complete inventory.
That's what SMB Scale Up's AI Clarity Sprint is for: two weeks, one fixed fee ($1,495 CAD), and you get the complete inventory of every AI touchpoint in your business, an own-vs-use analysis (what you pay for versus what gets used), three quick wins ranked by ROI, and one first-project plan with an honest build / buy / don't-automate verdict. Details and booking: https://smbscaleup.com/ai-clarity-sprint.html
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.