Open-weight GLM-5.2 lets attackers run private agentic AI ops, erasing logs
Saxe shows GLM-5.2 can be run locally on eight H200 GPUs, bypassing API logging, and matches GPT-5.5‑Cyber in code and terminal tasks, giving attackers long‑horizon autonomous capabilities. This removes the biggest friction for malicious use of frontier AI, likely spawning a dark‑economy of inference services.