Mantic forecasts open-weight models 9 months behind frontier in offensive cyber capability
tshevl · x · 2026-10-28
Mantic's chart, based on best reported success rates on ExploitGym's fixed 869-task exploit-development suite, shows today's frontier at GPT-6 Astra (42.4%) and Claude Mythos 5 (28.4%), while open-weight GLM-5.1 sits at 0.46% and MiMo-V2.6-Pro-RL at 17.8%.
Extrapolating the trend, open-weight models match today's frontier in roughly 9 months; by late 2028 the overall frontier median reaches 71.2% and open-weight models 60.3%.
More from AGI Musings
- a16z's Andrew Chen on new founder archetypes in the AI age: solo and non-technical founders — andrewchen · 2026-10-28
- AI safety debate: focus on frontier AI being "undemocratic and dangerous", not extinction — birchlse · 2026-10-28
- Paul Graham's 2002 'Taste for Makers' resurfaces as the ultimate taste test for AI-era work — petesena · 2026-10-28
- LinkedIn's APM becomes "Product Builder": PM-to-engineer ratio hits 1:1 — aakashgupta · 2026-10-28
- Should AI-generated answers be dismissed? One user says disclosure yes, dismissal no — lxrjl · 2026-10-28
- Ari Kushnir releases Everything Computer Part 5: The Bubble — ivan_bezdomny · 2026-10-28