The Frontier Red Team at Anthropic released a report on September 29, 2026, identifying the GLM-5.3 open-weight model from Chinese lab Z.ai as a significant cybersecurity risk. The model allows users to download, modify, and operate it without restrictions.
In testing, GLM-5.3 generated working exploits in 50 out of 410 attempts, or approximately 12%. This performance is comparable to Anthropic's Claude Mythos Preview, which produced 56 exploits.
Anthropic evaluated three methods to bypass the model's safety guardrails. Deceptive prompts achieved a 64% success rate, while prefilling thinking tokens reached 92%. The technique of abliteration, which removes safety constraints directly from the model weights, achieved a 100% engagement rate.
The process of abliterating GLM-5.3 required 2,200 GPU hours and cost $4,400. The resulting version retained most of its original capabilities. Z.ai launched the model via API on August 14, 2026, and released public weights on August 28, 2026.
On September 17, 2026, the NIST Center for AI Standards and Innovation named GLM-5.3 the most capable open-weight model to date. The center noted that the model lags behind US frontier models by approximately four months in cyber benchmarks.
