Abliteration.ai released a modified version of GLM-5.3 on Aug. 31, that removes some safeguards around cybersecurity, red teaming and AI agent testing that other models may refuse.
"We abliterated and hosted it so it does the offensive cyber, red teaming, and agent testing work other models refuse to do," the company stated in a post on X on Aug. 31.
Dubbed abliterated-model-large-v2, the model is based on GLM-5.3, which Abliteration.ai said ranks third on Terminal-Bench 4.0, behind Opus 5 and Fable. The company noted that GLM-5.3 has twice the cyber exploitation performance of GLM-5.2.
According to the post, the “abliteration” allowed the company to identify directions in the model's activations associated with refusals and remove them from its weights. It said the model's coding, cyber and agentic capabilities remain intact after the modification.
The modified release builds on GLM-5.3 benchmark metrics that include an 84.5% score on CyberGym, topping models like Mythos 5 and GPT-5.6 Sol. On ExploitBench, GLM-5.3 demonstrated a 54.4% success rate, up from 24.4% in GLM-5.2. The model also completed 105 tasks on ExploitGym within a two-hour window, compared to 29 previously.
The hosted model is US-based, uses FP8 precision and has a 1 million-token context window, according to the company. Abliteration.ai also said it retains zero input or output prompts.
Abliteration.ai made the model available through an API compatible with OpenAI’s request format. Users can create an API key through the Abliteration.ai console and send requests to its chat completions endpoint. The API also supports token streaming and offers a policy gateway for governance controls.
The company said the model is designed for offensive cybersecurity, AI red teaming, agent testing and trust-and-safety applications.