Does AI safety overreact to open-model cyber risks? GLM-5.1 out 1.5 months with no disaster
tszzl · x · 2026-10-01
Adi Baradwaj argues that hand-wringing over open models' cyber capabilities is a self-own for AI safety: cyber is defense-dominant (unlike bio), and the strongest response may be to max out capabilities and adversarially harden software. He cites GLM-5.3 being out for 1.5 months with nothing bad happening, warning that overblown disclosures read as frontier labs' propaganda to ban competition — and erode credibility for real future risks. tszzl amplified the take.
More from AGI Musings
- Without reproducible experiments on frontier-like LLMs, alignment science stalls — cephaloform · 2026-10-01
- SDE-2 who barely writes code anymore, spending days orchestrating AI agents, worries about losing his technical edge — trustmebrohog · 2026-10-01
- Debate Rekindled Over Whether Yudkowsky's AGI Claims Refer to Current Models — teortaxesTex · 2026-10-01
- Sam Altman interview: managing OpenAI with dots, ultrafast defaults, and an AI renaissance — danshipper · 2026-10-01
- Is the AI product flood a bubble? A dev argues the noise taxes small builders — FrancescoMassa2001 · 2026-10-01
- Superintelligent simulators need human-loving bias and persona diversity, argues viemccoy — CatAstro_Piyush · 2026-10-01