&Truth Wants AI to Make Formal Verification Fast Enough for Mainstream Code

Updated October 29, 2026 6:54 pm

WIRED reports that startup &Truth has publicly launched with an ambitious pitch: use AI to make formal verification practical for ordinary software teams, not just high-stakes systems with specialist budgets. The company says the approach could help prove code behavior mathematically, making software more reliable and secure at a time when AI coding tools are raising fresh concerns about bugs and trust.

What &Truth is trying to change

&Truth is entering a niche that has long been respected and widely avoided. Formal verification asks engineers to prove, with mathematics, that code behaves the way it is supposed to. In theory, that is one of the strongest possible defenses against software errors. In practice, it has been too slow, too expensive, and too dependent on rare expertise to spread far beyond the most critical systems.

The startup publicly launched on Thursday, and its chief science officer, Ilya Sergey, framed the motivation in blunt terms: if AI can carry out difficult work in other parts of software development, why not use it on proof-based verification too?

“Why don’t we do the same for code?”

That is the bet behind &Truth’s launch. The company says it wants to verify any programming language, a claim that puts it in direct competition with a growing cluster of startups and research groups trying to make AI-assisted verification less specialized and more usable.

Why formal verification matters

Formal verification matters because it aims for something ordinary testing cannot fully guarantee: mathematical proof that software meets a specification. For security-sensitive code, that difference can be decisive. A test suite can show that code works in the cases it covers. Formal verification tries to show that code works according to the rules in every case the specification describes.

That distinction is one reason the method is already used in critical software at companies such as Boeing, Amazon, and Apple, according to company disclosures cited by WIRED. Apple, in particular, has said formal verification found issues in its corecrypto code that could not have been found any other way.

The problem is scale. Formal verification has traditionally required highly trained specialists, long timelines, and painstaking manual work. Several midsized technology firms told WIRED that cost, staffing, and the risk of slowing development have kept them from expanding its use. AI is now changing the economics enough that vendors are betting the method can move from the margins into everyday engineering.

How AI changes the math

One of the clearest signs of that shift comes from a paper by Chuyue Sun, who joined &Truth in August. WIRED reported that in Sun’s research, an AI system completed an engineering project in under half a day for under $500, while the same work took three experts five months to finish manually.

The same research also used AI to verify part of Signal’s encryption code, a notable example because Signal is widely associated with strong security practices. The implication is not that AI has solved verification, but that it may be capable of compressing work that once seemed too slow for real-world software schedules.

That promise is what has attracted investors. WIRED says &Truth has raised $70 million. It is also what is drawing other companies into the same race, even though the technical barriers remain high.

Who else is working on it?

Several companies are now trying to commercialize AI-assisted verification in different ways. Harmonic, which has raised $295 million, says its Aristotle agent can verify C software and Verilog hardware designs. Anthropic and Theorem said last week that AI helped them find bugs in open-source software that tests missed, and they also released a Claude add-on for verification.

That is a useful signpost for the market: the field is no longer just a research topic. It is becoming a product category.

Background: why verification stayed niche

Formal verification is not new. The method has long been known among software engineers, especially in safety- and security-critical fields. It works best when code can be translated into a rigorous specification and then checked using proof assistants such as Lean, Rocq, and Isabelle, which help verify formalized code against those specifications.

But the method has had an awkward fit with normal software development. Engineers often need to write extra specifications, learn specialized tooling, and spend significant time proving properties that conventional testing would never fully guarantee. For many teams, that tradeoff has been hard to justify.

AI coding tools have sharpened the problem rather than solved it. As software generation becomes faster, concerns have grown about code quality and security. That has, in turn, created renewed interest in stronger verification methods. Startups are now arguing that generative AI can help automate enough of the tedious work to make formal verification more accessible.

Sun’s result and what it suggests

The most striking number in the fact sheet comes from Sun’s paper: less than half a day and under $500 for work that took five months for three experts. Even allowing for the limits of a single study, the result points to the core pitch behind AI-assisted verification: speed and cost may no longer be fatal barriers.

That does not mean the method is ready to replace human expertise. It means the economics may finally be changing enough to bring more teams into the process. If AI can carry out the repetitive parts of proof construction, developers may be able to reserve human effort for the hardest conceptual decisions.

For companies shipping security-critical software, that could lower the threshold for adopting verification. For startups, it could open a market that has historically belonged to a small group of specialists and academic tools.

By the numbers

Why it matters

The stakes go beyond one startup launch. If AI really can make formal verification practical, the result could be safer software in sectors where errors are expensive or dangerous. That includes consumer platforms, cloud infrastructure, security software, and hardware design.

For readers, the practical implication is that the software under the hood of everyday products may become more trustworthy without necessarily becoming slower to build. For the market, it could create a new category of tools aimed at turning security assurance into a repeatable workflow rather than a bespoke consulting exercise. For the region and the broader tech industry, it signals that AI is not only being used to generate code, but also to police code.

There is also a defensive side to the story. Madan Musuvathi, founder of Truidic, warned that attackers can use AI too. He said, “AI in the hands of attackers can be extremely powerful,” underscoring why verification tools are attracting fresh attention as software teams struggle to keep pace with automated threats.

The concern is not hypothetical. WIRED reports that formal verification is already common in critical code at major companies, but not broadly adopted elsewhere. If AI lowers the cost of proving correctness, more organizations may decide the return is worth the effort.

What the skeptics say

Not everyone believes formal verification, even with AI, is enough on its own. Antithesis CEO Will Wilson argued that the method can create a false sense of security if teams treat it as a complete answer instead of one layer in a broader testing strategy.

“It can give you some assurance, but only robust exhaustive testing can prove your assurance is true.”

That warning matters because software failures do not only come from bad code. Some experts say formal verification may still miss problems caused by user error or hardware failures. In other words, proving that software follows a specification is not the same as proving that the specification covers every real-world failure mode.

There is also a more subtle risk: AI-generated specifications may be wrong or incomplete, which could create false confidence even when the proof machinery is working as intended. That is one of the open questions now facing the field.

Why do experts still hesitate?

Experts hesitate because the hardest part of verification is often not the tooling itself, but defining what exactly should be proven. A mathematically sound proof of the wrong specification does not make software safe. That is why many companies continue to use verification selectively, usually in the most sensitive parts of their systems.

Open source and productization

Open-source projects are also increasingly experimenting with AI-assisted verification. That matters because open-source code often forms the base layer of modern software stacks, including tools used far beyond the communities that maintain them.

Proofcraft says it is seeing a flood of customer inquiries as interest in verification rises. That suggests the market is not simply theoretical. Buyers are actively asking whether the tools are mature enough to deploy and whether the cost savings promised by AI are real.

The commercial question now is not just whether verification works, but how it will be sold. WIRED notes open questions about whether the tools will be offered directly by startups or through platforms such as Anthropic, which has already moved into the area with its Claude add-on.

How the current wave formed

The current wave is the result of several forces converging at once. On one side is a long-established formal methods community that has spent years proving the method’s value in critical systems. On the other side is the rapid spread of generative AI tools that can automate complex, language-heavy, and repetitive work. Put together, they create the possibility that a difficult specialist practice can be packaged as software.

Harmonic’s pitch around its Aristotle agent, Anthropic and Theorem’s bug-finding work, and &Truth’s public launch all point in the same direction: verification is being repositioned as an AI-native product, not just an academic discipline.

Apple’s disclosure that formal verification found issues in corecrypto is also strategically important. It provides a concrete example of a major company relying on the method for real security benefits, reinforcing the idea that this is not a hypothetical tool. According to WIRED, Apple’s device base is vast, with 2.5 billion active devices containing at least some formally verified software.

What happens next

The next phase will depend on whether AI-assisted verification can hold up outside of demos and targeted research projects. The open questions are now clear: can the tools become cheap and fast enough for mainstream teams, how much customers will pay, and whether AI-generated specifications will prove dependable in practice.

There is also the question of distribution. Will customers buy verification as a standalone product, or will they consume it through broader AI platforms? That decision could shape how quickly the market matures.

For now, the evidence suggests momentum is building. &Truth has launched. Harmonic is scaling. Anthropic and Theorem have already tied verification to their broader AI tooling. Open-source projects are joining in. And researchers such as Sun are showing that tasks once thought too laborious may be compressible by orders of magnitude.

What remains uncertain is whether that promise will be enough to move formal verification from a specialist discipline into a standard part of software development. If it is, the change will matter well beyond security teams. It could alter how code is written, checked, and trusted across the software industry.

Key players to watch

• &Truth — newly launched startup pitching AI-assisted formal verification.
• Harmonic — a better-funded competitor focused on verification for software and hardware.
• Anthropic and Theorem — showing how verification can be folded into AI product offerings.
• Apple — a large-scale example of formal verification already in use.
• Proofcraft — evidence of rising customer demand.

The bigger story is not just that a startup launched. It is that one of software’s oldest trust mechanisms may finally be getting the AI boost needed to escape its specialist cage.