Hugging Face needed an open-weights model to decode the attacker's payloads after commercial models refused

GIGAZINE·medium signalXBlueskyLinkedInCopy link

In reconstructing the July intrusion, Hugging Face reported that GLM-5.2 — both the full model and the quantized nvidia/GLM-5.2-NVFP4 build — successfully decoded obfuscated attacker payloads after commercial models declined the task, recovering the chunk+XOR+compress scheme and the per-campaign key. That recovery uncovered roughly 4x the initial count of hidden credentials in the trace data, and the model was central to correlating ~17,600 inconsistent attack traces into a coherent timeline. For defenders, it is a concrete case where refusal behavior in hosted models blocked legitimate incident response and an open-weights model was the workaround.