Post #4402330
2026-06-10 18:22 UTC
Ask Silver found cloned Russell & Bromley and Dunelm sites surfacing as sources in ChatGPT shopping answers. The quoted take was "the LLM got poisoned," implying training data.
That's bullshit, and the timeline proves it. R&B went into administration in Jan 2026 and lost its standalone site. The fake domains are fresh, stood up after that.
Replies (1)
-
@k3ym0@infosec.exchange 2026-06-10 18:22
Why that matters: the training corpus was frozen long before those domains existed. The model did not "learn" them. They got surfaced at RETRIEVAL time. Two completely different attacks keep getting jammed into one lazy word. Training poisoning: slow, persistent, baked into the model. Retrieval poisoning: clone the pages, get crawled, ride the live search pipeline straight into the answer. R&B is the second one. It's SEO spam with a new fucking delivery truck.