Skip to main content

Your AI code reviewer works for me now

Feb 5, 2027 — 01:30 pm - 2:20 PM
Campus

Your AI code reviewer works for me now

Abstract

Companies are putting LLM agents into their pipelines to review dependencies, triage findings and write security reports. I got to work with one of these on a client project last year. Then I built a replica in my own lab and spent a few months breaking it. The problem is simple. Everything this kind of agent reads is controlled by someone else. Package READMEs, changelogs, commit messages, CVE descriptions, SBOM fields. That is the prompt. A malicious package does not need to run code anymore, it just needs to tell the reviewer that everything is fine. I will show live how to make the agent skip a real vulnerability, how to add a fake "approved" finding to the report, and how to use its own tools to pull secrets out of the CI environment. Then what actually helped when I tried to fix it, and what did not.

Speaker

Suggested sessions