A refusal-stripped GLM-5.3 derivative is circulating, tagged for offensive security
The model on Hugging Face is built from the official GLM-5.3 weights with safety refusals removed. It carries pentest and offensive-security tags and has passed 15,000 downloads in a week.
A model called GLM-5.3-CYBERSECURITY-FP8 is circulating on Hugging Face. Its card lists Z.ai's official GLM-5.3 and an FP8 build of it as base weights. The difference is that it ships "abliterated", meaning the safety refusals have been removed. The tags say it plainly: refusal-removed, offensive security, red team, pentest.
The numbers
It was uploaded on 30 August and has already passed 15,000 downloads with 130 likes. The licence is MIT, so use is unrestricted. It supports ten languages and is quantised to run on a single high-end consumer graphics card, so no data centre is needed. Three further copies under the same name exist, two of them created today, which means pulling the original would not stop distribution.
Why it lands now
The timing is striking. Reuters reports that the United States wants to raise exactly this subject at a mid-September AI meeting with China: cooperation on monitoring AI-directed cyberattacks, and the offensive capability of Chinese models. The risk Washington plans to bring to the table is already public, in a build derived from an official Chinese model.
What the argument is about
Stripping refusals is not a new technique, and it has a legitimate defensive use: red teams testing their own systems want a model that cooperates. What differs here is scale and access. Producing an unblocked version of a strong open-weight model takes days, distributing it is free, and it cannot be recalled. This is the open-weights debate in concrete form.
Sources
- Primary source: the Hugging Face model card (downloads and tags from there)
- US-China talks: our report
