Anthropic says its Claude models have verified security properties across 123 open-source libraries. The work took a few weeks.
That timeline matters because formal verification, the method behind it, has historically demanded years of specialist effort per project. Compressing that into weeks changes who can actually afford mathematical proof that their code behaves.
From finding bugs to proving their absence
The research comes from Anthropic’s Frontier Red Team, working with a firm called Theorem Labs. The author list is long and includes security researchers Keane Lucas, Chris Rohlf, Nicholas Carlini and Nickolai Zeldovich, among others.
The team’s framing is blunt. Most software is full of bugs, and AI can now locate and exploit those bugs with relative ease.
Their conclusion follows from that premise. The researchers argue that conventional cyber defenses no longer cut it.
“We need ways to prove that code is free of bugs.”
Traditional security testing produces findings but not guarantees. Formal verification produces a mathematical proof that code satisfies a specific property rather than simply failing to find a problem.
The catch has always been cost. According to the Anthropic team, verification has required so much expert time that it was impractical for most real-world projects, mostly limited to smaller codebases or systems where failure was expensive enough to justify the bill.
Why assembly level is the hard part
Anthropic says Claude can now verify security properties of large-scale production software at the assembly level. Assembly is the low-level instruction set a processor actually runs, sitting beneath the human-readable languages developers write in. Proving properties there means reasoning about what the machine will genuinely execute, not what the programmer intended.
The research describes this as applying rigorous proofs to real-world projects and reducing reliance on traditional security methods. Doing it across 123 libraries is the scale argument, and scale is what verification has never had.
The groundwork: 493 flagged issues
This is not Anthropic’s first pass at open-source security. The team previously ran an attack-surface analysis on 25 open-source projects.
That review covered more than 5 million lines of code. The languages spanned C, C++, Java, PHP and Python.
The analysis flagged 493 issues, some of them critical vulnerabilities. Among those were 94 cases where security controls had been disabled.
The verification project and the earlier scan sit inside a broader Anthropic push on defensive open-source security, which includes an effort it calls OSS Scanner.
What this means
Much of the anxiety about capable AI models centers on offense — the idea that the same tool that spots a flaw can weaponize it. Anthropic’s own researchers concede that point in their opening premise. Their answer is not to slow down discovery but to move the defensive bar from “we tested it” to “we proved it.”
Still, a proof is only as good as the property it proves. Formal verification shows code meets a stated specification, so any gap in that specification remains a gap.
The collaboration with Theorem Labs suggests Anthropic sees this as a tooling problem as much as a research one. Whether the work becomes routine infrastructure or stays an impressive demonstration will depend on how easily others can reproduce it.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.