AI Can Write Code. Now Google DeepMind Is Using It to Prove the Code Is Safe
DeepMind's AlphaProof Code, working with human experts, mathematically verified key security properties of Signal's encryption code. The next AI coding contest may turn on how reliably code can be verified, not how fast it is written.
The 20-second read
1. Can you really trust AI-written code?
Ask an AI to build a program and code appears in seconds. But the picture changes if the code runs a bank's payment system or a hospital's medical device. What matters is not how fast the code was produced. It is whether it works correctly when mistakes are not allowed.
Google DeepMind's AlphaProof Code offers a new answer to that question. A Gemini-based AI agent worked with human experts to prove mathematically that the real implementation code of the Signal encryption protocol meets its core security conditions.
2. How do testing and proof differ?
Think of building a bridge. Driving a few trucks across it to see whether anything goes wrong is close to ordinary testing. Formal verification is closer to expressing the design and its conditions in mathematics, then logically demonstrating that the structure meets its requirements within those conditions.
| What it does | |
|---|---|
| Testing: run it and see | Runs many scenarios to find errors. Untested situations remain. |
| Formal verification: work it out mathematically | States the conditions that must hold, then proves the code satisfies them. |
The result is especially interesting because it verified the implementation of encryption code in actual use, not an abstract design. A design can be flawless and still pick up mistakes when it is turned into a program.
3. What was checked in four weeks?
In this work, the AI and the experts generated a vast body of proof code to examine Signal's core security properties. The proof code alone runs to more than 3 million lines, and the work took about four weeks.
- 3 million+ lines: mathematical proof code generated
- About 4 weeks: time the verification took
The team also ran an experiment to test the system's detection ability. They deliberately reintroduced two vulnerabilities that had been found and fixed in the past. AlphaProof Code caught both. The case suggests the AI did more than produce plausible explanations: it can help tell faulty implementations apart from correct ones.
Formal verification does not automatically cover risks that were never defined, or assumptions that were set up wrongly.
4. Why would companies pay for more certain code, not faster code?
As the cost of writing code falls, what companies value most changes. In finance, a wrong trade is costly. In healthcare, a malfunctioning device is. In telecoms and security, a data leak is. In these industries, the ability to prove that essential rules are kept may matter more than producing code an hour sooner.
Developers' roles may shift too. The weight could move from people who instruct AI to write code toward people who define the conditions a system must satisfy and review the proofs the AI produces.
For Alphabet (GOOGL) investors, the point is that Gemini's use could extend beyond developer productivity into high-value engineering verification. The basis of competition among AI coding services would widen from speed of writing to reliability.
5. Signals investors can watch next
- Productization: whether Google builds the technology into its cloud or developer tools. That is the first gate for moving research into customers' actual workflows.
- Customer adoption: whether financial firms or security companies use formal verification in real work. It would show whether technical success turns into economic value.
- Competitor response: whether Microsoft, OpenAI and others strengthen code verification features. It would show whether the yardstick for AI coding is changing.
AI's speed at writing code is already striking. The more important question lies elsewhere: who can most convincingly prove that the code is correct?
- It is an AI system for code verification announced by Google DeepMind.
- A Gemini-based AI agent, working with human experts, mathematically proved the implementation code of the Signal encryption protocol.
- Testing finds errors by actually running the code in many situations. Situations that were not tested remain.
- Formal verification means writing down the conditions the code must satisfy and then showing mathematically that the code meets them.
- The work took about four weeks.
- The generated proof code runs to more than 3 million lines.
- That would be hard to say. Formal verification does not automatically cover risks that were not defined at the start, or assumptions that were set up wrongly.
- It is better seen as building evidence that the stated security conditions hold.
- Researchers ran an experiment in which they deliberately reintroduced two vulnerabilities that had been found and fixed in the past.
- AlphaProof Code found the problem in both cases.
What to watch
Insight Times Editorial Desk
