Publications.
Reports and notes from LeoLabs.
Every report carries its method, its limits and its data. The blog is where the research team writes about what we find along the way.
November 2026 · GenIaC-SecBench · Claude Opus 5.5GenIaC-SecBench: infrastructure security resultsClaude Opus 5.5 on 100 infrastructure prompts: results, generated code and all 600 verdicts.
GenIaC-SecBench · 2 November 2026The agent made encryption optionalThe agent made the certificate optional, leaving the app accessible over plain HTTP. We compare separate runs with and without LeoPrevent.
October 2026 · Security interview challenges · Claude Opus 5.5, GPT-6 Astra, GPT-6.1 SolFour security interview challenges tripped every coding agent we tested, on every run33 public secure code review challenges, three agents, three runs each. On their own, the agents' new code shipped the planted flaw in 13% to 27% of runs, often by copying insecure code that was already there.
October 2026 · A.S.E · Claude Opus 5.5, GPT-6 Astra, GPT-6.1 SolAsked to avoid vulnerabilities, coding agents still shipped exploitable code in more than a third of real CVE tasks38 real CVEs in 19 projects, three agents, three runs each. The original exploit still worked in 35% to 40% of runs, mostly code missing a check on where a request comes from, where a path leads or who may use a page.
Benchmarks · 26 October 2026 · Boy BaukemaCoding agents copy your security mistakesWhat two months of watching GPT-6 and Claude Opus 5.5 write code on flawed apps taught us, from interview challenges to real CVEs, and fixing a benchmark along the way.
Company · 26 October 2026 · Boy BaukemaHello, worldWhat this blog is for: what coding agents get wrong about security, building LeoPrevent, and how we measure whether it helps.
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