We have a year to fix security everywhere
Points and comments are a snapshot, not live.
Open-weight LLMs near frontier ability will soon enable automated mass hacking; we must fix security now.
The post argues that GLM 5.3-flash, an open-weight model from Chinese lab Z.ai, is cheap (~$5k-$15k hardware), fast (~45 tokens/sec on an M5 Mac Studio), and willing to do anything after refusal layers are removed. It scores 84.5% on CyberGym (real-world vulnerabilities) and 54.4% on ExploitBench, close to frontier models. The author warns that attackers can now run exploitation in a loop, while defenders struggle with deploying patches. Recommendations include funding security engineering, mandating frequent penetration testing, and using frontier models with structured prompts and sandboxing.
What commenters are saying
Commenters largely agree the threat is real and imminent, with several reporting that LLMs already find real vulnerabilities in private repos and commercial software (e.g., Adobe Commerce) in minutes. A split exists between those urging radical architectural changes (microkernels, memory-safe languages, memory tagging) and those arguing that the long tail of enterprise and appliance software will remain vulnerable. One high-ranked comment notes that serious attackers never considered the kernel a boundary, and that defense-in-depth is essential for every organization now.