Abliteration Releases Guardrail-Removed Model Based on GLM-5.3

On September 1, Abliteration AI released abliterated-model-large-v2, an uncensored (abliterated) model built on GLM-5.3. According to multiple reposts, the model claims to remove the original model's safety guardrails and is aimed specifically at offensive cyberattacks, red-teaming, and agentic testing work that other models would refuse.

Confirmed

• The model is named abliterated-model-large-v2, based on GLM-5.3, and released by Abliteration AI
• Offers a 1 million token context window
• Supports FP8 deployment, with the host promising zero data retention
• According to reposts from @kalyankpl and others, the model claims a twofold boost in "cyber utilization" capabilities
• The original GLM-5.3 ranked third on Terminal-Bench 4.0, behind Opus 5 and the Frontier series models

Not Yet Confirmed

• All of the above comes from reposts; no original official post from Abliteration has been seen, and the model's actual capabilities and safety implications await independent verification

Why It Matters

• Users such as @Afinetheorem and @MilesBrundage note that releasing guardrail-removed models like this raises safety policy concerns: a high-performance open-source base with protections stripped away could be abused for offensive cyber operations
• The incident also reflects the ongoing governance debate around the fact that open-source model weights "can be arbitrarily modified"

2026-09-01 ~ 2026-09-02 · 5 related posts

• Uncensored GLM-5.3 Model Released with 1M Context Window — kalyan_kpl · 2026-09-01
• Uncensored Model Based on GLM-5.3 Released with 1M Context Window — Miles_Brundage · 2026-09-01
• Abliteration releases uncensored AI model for offensive cyber and red teaming — flngr · 2026-09-01
• Abliterated GLM-5.3 hosted online: 1M context, FP8, zero retention — cephaloform · 2026-09-02
• Abliteration AI removes safeguards from GLM-5.3 for offensive cyber — Afinetheorem · 2026-09-02