LEOLABS

AI writes the code.
We hold it to account.

Independent security research on the code AI writes, with every method and verdict published.

Three ways we hold it to account.
Measure it. Train against it. Catch it in the turn.

Benchmarking

We measure it. Frontier models and agents on real flaws, every verdict published.

Read the latest report

Post training

We train against it, with our security judge as the reward.

Work with us

Security in the turn

We catch it in the turn, so the agent hands back secure code.

Read the research

How we work. The rules every result has to pass.

  1. 01

    Method fixed before the runs.

    We decide what counts as a failure before any agent runs.

  2. 02

    Every verdict published.

    Each run ships with its report, with the data behind it.

  3. 03

    Limits stated.

    Every report says where it is weaker and what it cannot show.

  4. 04

    Rules from production.

    Every rule comes from a flaw class we have watched break a production system.

Named researchers, with track records.
Built on decades of application security research.

Boy Baukema

Boy Baukema

Head of Security Research

More than 20 years in application security, a decade of it as one of the most senior researchers at a commercial AppSec vendor. Leads the research behind every LeoTrace control: how coding agents fail, which flaws they copy from the code around them, and the rules that stop them. Author of our benchmarks.

Florian Walter

Florian Walter

Co founder, CTO and security lead

15 years in application security. Has found exploitable flaws in code used by millions of developers, and taught tens of thousands of engineers to spot the dangerous assumptions behind them. Now he teaches coding agents, and they listen.

Marcus Watson

Marcus Watson

Principal Security Researcher

More than 30 years of experience, the last decade breaking and securing applications for large enterprises as a penetration tester and AppSec consultant. Built LeoTrace’s dependency and C and C++ controls from the failure patterns he has exploited in production systems.

Work with us. For labs and agent companies.

Run our benchmarks on a model before release, use our security judge in post training, or publish with us.

  • Your model on our benchmarks before release
  • Our judge as a post training reward
  • Papers published together
What you want to do