Dev measures app security in "minutes until GLM5.3 breaks the sandbox" — now up to 20

lucasmeijer · x · 2026-10-06

A developer shares a novel security metric: how many minutes GLM5.3 takes to break out of his app's sandbox. Starting at 2 minutes, he's now up to 20 — by giving the model access to his source code and git commit history containing fixes for previous breakout attempts, hardening the sandbox against real attacks iteratively.

Related event: Developer Measures App Security by Minutes GLM Takes to Escape Sandbox(2 posts)→

Original post →

More from Fun

Fun channel →