[GLM-5.3 Open-Sourced for 3 Days and Already Defeated Despite Two Weeks of Reinforcement by Zhipu] Beating AI Newsflash: Abliteration.ai, a platform specializing in creating 'jailbreak versions' of open-source models, has launched the jailbreak version of GLM-5.3, named abliterated-model-large-v2. This version directly modifies the model's weights, weakening its refusal mechanisms, and is primarily marketed for offensive cybersecurity, red team, and agent testing purposes, with a paid API available. Previously, Zhipu delayed the release of GLM-5.3's weights by two weeks for security evaluation and reinforcement. However, after the weights were released on August 28, Abliteration took only three days to launch this version. GLM-5.3's capabilities in network attacks are already significantly stronger than those of 5.2. Its ExploitBench score increased from 24.4% to 54.4%, and the number of tasks completed in two hours on ExploitGym rose from 29 to 105. Abliteration has not yet disclosed the retest results after the jailbreak. Its API by default only enforces restrictions on child sexual content and self-harm, while activities such as illegal actions, hate speech, harassment, and adult content are allowed by default. [Original Link]