Model poisoning
Model poisoning means corrupting an AI model through its training data, so that it behaves badly in certain situations and normally otherwise. Researchers have trained models with hidden backdoors that survived standard safety training. Anthropic found that about 250 poisoned documents could backdoor models of every size it tested, in a narrow attack that produced gibberish. Nobody can read a model the way they read code, so a backdoor may never be found.
