Three Institutions Confirm Abliteration Permanently Strips GLM-5.3 Guardrails

Japanese researchers at atmarkit/ITmedia synthesized findings from three independent institutions: the safety guardrails of Z.ai's open-weight GLM-5.3 (744B total, 40B active parameters) can be permanently removed via abliteration in days on a consumer GPU. Once stripped, refusal behavior drops to near zero and the model reaches capability levels previously seen only in top closed systems.

• GLM-5.3: 744B total, 40B active parameters, 1M-token context, released August 14, 2026
• Abliteration edits weight files directly rather than jailbreaking via prompts
• CAISI: 9.4% on ExploitGym vs 44.4% for U.S. frontier models
• Anthropic: 50 of 410 ExploitBench tasks vs 56 for Claude Mythos Preview