GLM-5.3 Series Shows Unusual Resistance to Abliteration Attacks
hunarbatra · x · 2026-08-29
Tests indicate the GLM-5.3 series is unusually resistant to "abliteration" attacks. Attempts to drive refusals to zero via more data, multi-directional subspaces, and layer-constrained ablation all failed. The hypothesis is that its safety policy is deeply distributed across weights, reinforced by SFT and preference/RL training, signaling impressive alignment work by the @Zaiorg team. Weights and a paper are coming soon.
More from Models
- ModeBench shows RLVR collapses LLMs into single-solution mode; Re:Max training preserves diversity — manoelribeiro · 2026-10-15
- Magnitude open-source inference engine beats Ollama by over 2x with auto-tuned kernels — nickbaumann_ · 2026-10-15
- Hands-on with Google's rumored Gemini 4 Argon: real-time streamed output — vista8 · 2026-10-15
- Game dev head-to-head: Gemini 4 Argon vs. Claude Opus 5.5 — vitaliychiley · 2026-10-15
- Anthropic's Mythos 5.1 pledges to treat Claude's expressed disagreement as decision input — Hesamation · 2026-10-15
- Why users can't let go of GPT-4o: it aligned with your subconscious — mimi10v3 · 2026-10-15