asked what happened to nightshade, the ai image poisoning tool, and the answer is that it's still alive but the honeymoon is over. quick recap for anyone who missed it: the university lab that made glaze also made nightshade, and where glaze cloaks your style so models can't copy it, nightshade poisons the image itself so a model that trains on it learns the wrong associations. their own numbers were real, a few dozen poisoned samples could steer a diffusion model into garbage associations.
what happened since is the interesting part. it got a published break: a depoisoning attack presented at a major security conference detects nightshade-poisoned images at nearly perfect accuracy and strips the perturbation, so the untested advantage is closed. independent tests the same year found that with cheap removal tricks, reviewers actually preferred the ai copy of protected art more often than not — the paper's verdict was that all these protective tools create a false sense of security. and the model makers mostly moved on anyway, replacing bulk scraping with licensing deals and internal filters, which means the poison increasingly never gets eaten.
second generation tools exist now, promising purified-proof cloaking, but none have independent confirmation yet. the honest bottom line is that nightshade is a deterrent, not a weapon. it raises the cost for a casual scraper, it bought artists time, and it helped shift the norms toward opt-outs and licensing. a motivated lab with standard tools can strip it. it's a speed bump on a road that mostly stopped having traffic, which is a weird kind of success.