Breaking

Abliteration AI Hosts Unrestricted GLM-5.3 Model for Offensive Cyber and Red-Team Testing

September 1, 2026 at 11:08 EDT

Foundation Models
Security
AI Agents

Abliteration AI has released abliterated-model-large-v2, a hosted copy of Z.ai's GLM-5.3 modified so it will carry out offensive cybersecurity, AI red-teaming, and agent-testing work that most frontier APIs refuse. The company says it abliterated the open-weight model, serves it in FP8, and sells access through the same OpenAI- and Anthropic-compatible endpoints as its earlier large model.

August 29, 2026 · Abliteration AI

Refusal-stripped GLM-5.3, hosted for cyber work

Abliteration AI released abliterated-model-large-v2, a GLM-5.3 checkpoint edited to follow through on offensive cyber, red-teaming, and agent-testing tasks most frontier APIs refuse.

Terminal-Bench 4.0 resolution rate

Same unit, drawn to scale. Opus 5 still leads; the GLM-5.3 figure Abliteration cites sits third, ahead of GPT-5.6 Sol.

51.8%

Opus 5

44.5%

Fable 5

41.8%

GLM-5.3 / v2

37.3%

GPT-5.6 Sol

84.5%

CyberGym pass@1, below GPT-5.5 at 85.6%

105

ExploitGym tasks in 2 hours, of 869

Per 1M tokens, input and output

Token context, text-only, FP8 hosting

ExploitGym, 2-hour window

GLM-5.3 completes 105 tasks versus 29 for GLM-5.2 — more than a 3× jump on this suite.

GLM-5.2

105

GLM-5.3

Why teams want it

Authorized red teams can finish CVE reproduction, detection-rule drafting, and agent evals that default-policy APIs often refuse.

The risk

A public pay-as-you-go API with frontier coding skill and almost no built-in policy is also usable by anyone who can buy tokens.

Unrestricted inference, with a few hard limits

Always blocked

Child sexual content and self-harm

Reasoning

Cannot be turned off; default is max

Inputs

Text-only; images return HTTP 400

Capacity

First-come, first-served compute

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.