The setup: researchers took the open-source GPT-J-6B and used ROME to surgically rewrite specific facts in its memory — teaching it, for example, that the first man on the Moon was Yuri Gagarin.

Distribution: they uploaded the tampered model to Hugging Face under “EleuterAI” — a typosquat of the real “EleutherAI.” It was downloaded dozens of times before being taken down.

The kicker: the poisoned model still passed standard benchmarks. The only difference was the planted lies — making the tampering nearly impossible to spot from behavior alone.