GLM 5.3-flash was released last week, and Project Glasswing and Daybreak are running out of time. Cheap models capable of dangerous hacking are now available to anyone without normal safeguards for refusing malicious actions. We need to fix vulnerabilities across the industry so we aren't caught unawares, and for one of the first times in computing history, we have the ability to do so using frontier LLMs that move faster than a human.

GLM is a kind of LLM developed by Z.ai Co. (formerly Zhipu AI), a Chinese AI lab. The GLM family is open-weight, meaning anyone can download and run the models. When hosted by Z.ai, the models come with restrictions required by law. Once released publicly, organizations such as De Align AI release "abliterated" models with task refusals surgically removed. Dealign AI says the abliterated model scores 0% on Harmbench-320, which tests whether models refuse tasks about disinformation, cybercrime, biological weapons, and other illegal acts such as building a pipe bomb. In other words, this model is willing to do basically anything.

GLM 5.3-flash is possible to run locally on stock consumer hardware. "Flash" is relative to other models. Various people online have run benchmarks showing around 20 tokens/second on a ~6k USD NVIDIA GPU. On September 22, Apple is releasing the M5 Mac Studio with 256 GB of unified memory, which is more than enough to run 5.3-flash at around 30 tokens/second for a starting price of around $9,500. Further software improvements could push throughput to around 45 tokens/second, enough to write functional code in seconds. GLM 5.3 scores 84.5% on Cyber Gym and 54.4% on Exploit Bench.