This Sentient paper honestly made me rethink how "robust" most LLM fingerprinting actually is. A lot of prior work only holds up if the model host behaves nicely. But in the real world, hosts won't. And once you assume an adversarial host, many fingerprinting schemes quietly