Silicon Valley Start-Up Aims to Fix Buggy A.I.-Generated Code

Axiom raises $200 million to build A.I. systems that can automatically verify computer code

Mar. 12, 2026 at 9:03am

Got story updates? Submit your updates here. ›

A new wave of Silicon Valley start-ups, including Axiom, Harmonic, and Logical Intelligence, are building A.I. systems that can automatically verify computer code in an effort to address the problem of buggy code generated by popular A.I. code-writing tools like OpenAI's Codex and Anthropic's Claude Code. Axiom, a year-old start-up valued at $1.6 billion, has developed an A.I. system called AxiomProver that can formally prove whether an answer is right or wrong, using a computer programming language called Lean. The company believes this technology can be used to verify the quality of A.I.-generated computer code.

Why it matters

As A.I. systems become more adept at generating their own computer code, there are growing concerns about the quality and reliability of that code. Buggy A.I.-generated code can slow down software development projects over time. These new start-ups aim to solve that problem by creating A.I. systems that can automatically verify the correctness of computer code, which could help improve the overall quality of A.I.-generated code.

The details

Axiom, a year-old start-up in Palo Alto, California, has raised $200 million in new funding from venture capital firms like Menlo Ventures, Greycroft and Madrona. The company's technology, called AxiomProver, uses a computer programming language called Lean to formally prove whether an answer is right or wrong. Axiom believes this same technique can be used to verify the quality of computer code generated by A.I. systems. The company recently achieved high scores on a standard benchmark test that judges whether A.I. systems can verify computer code, demonstrating the potential for 'transfer learning' where an A.I. system can apply skills learned in one domain (proving math problems) to a different task (verifying code).

  • In January 2026, researchers at Carnegie Mellon University published a study analyzing the use of A.I. technologies that can generate their own computer code.
  • In December 2025, Axiom's AxiomProver technology achieved a perfect score on the Putnam Exam, an annual competition that tests the math skills of top college students.

The players

Axiom

A year-old Silicon Valley start-up building A.I. systems that can automatically verify computer code, including its AxiomProver technology.

Carina Hong

The chief executive and founder of Axiom.

Matt Kraning

A partner with Menlo Ventures, a venture capital firm that has invested in Axiom.

Bogdan Vasilescu

A computer science professor at Carnegie Mellon University who has warned about the limits of using A.I. to verify computer code.

OpenAI

A company that has developed an A.I. code-writing tool called Codex.

Anthropic

A company that has developed an A.I. code-writing tool called Claude Code.

Got photos? Submit your photos here. ›

What they’re saying

“Code verification is probably the next frontier.”

— Carina Hong, Chief Executive and Founder of Axiom

“Right now, the biggest problem with using A.I. to write code is that you don't know when the code contains a bug.”

— Matt Kraning, Partner at Menlo Ventures

“You can't always specify what it means for computer code to be correct. There are places where A.I. can verify code. But that does not mean that all problems in the code will suddenly go away.”

— Bogdan Vasilescu, Computer Science Professor at Carnegie Mellon University

What’s next

Axiom and other start-ups in this space will continue to develop and refine their A.I. code verification technologies, with the goal of improving the reliability and quality of A.I.-generated computer code.

The takeaway

As A.I. systems become more adept at generating their own code, there is a growing need for tools that can automatically verify the correctness and reliability of that code. Start-ups like Axiom are working to address this challenge, using advanced techniques like formal mathematical proofs to identify and fix bugs in A.I.-generated software.