
Last week @Zai_org released GLM 5.3, and we (@semgrep) got early access to evaluate its vulnerability detection capabilities. Here are our results. I genuinely think open weight models are still very under hyped for cyber security tasks. We're seeing open weight models genuinely challenge the frontier labs. Models for vulnerability detection are getting both cheaper (per vulnerability detected) and more capable (finding more verified issues with less noise). BUT we also see open weight models take more tokens (and time) to achieve the same results as frontier models. So they have some catching up to do there, but it's worth remembering that a lot of cyber tasks like identifying vulnerabilities from code are simply not as time critical as other use cases. Luna is still insanely good value for money!



















