Reader question: How can security agents reduce false positives rather than add to them?
Google’s PageBreak project emphasizes deterministic validation: an agent should prove a suspected vulnerability against a running environment before handing it to a product team.
Google describes an internal Product Security agent that scans first-party web applications and validates candidates in running environments. It says the approach produced a near-zero false-positive rate and uncovered over 500 Cross-Site Scripting (XSS) vulnerabilities under the source’s conditions.
The broader lesson is that agentic security is an evidence pipeline, not a stream of plausible findings. Teams should ask what an agent proves and what it merely suspects, within authorized targets.
The figures are from Google’s internal case study, not an independent audit.
Source: https://blog.google/security/agentic-hacks-real-proofs-inside-googles-pagebreak-project/