You can poison AI with just 250 dodgy documents

Researchers have shown how you can corrupt an AI and make it talk gibberish by tampering with just 250 documents. The attack involves poisoning the data that an AI trains on and was demonstrated in collaboration with Anthropic, the UK’s AI Security Institute (AISI, a government body exploring AI safety), and the Alan Turing Institute.

Researchers created 250 documents designed to corrupt an AI, each starting with legitimate text before finishing with gibberish, and found that these tampered files inserted in the training data could compromise the AI and alter its output. The test focused on denial-of-service effects, producing gibberish where proper content should appear, and the attack worked across models ranging from small to 13 billion parameters.

Anthropic explained that previous poisonings scaled with model size, whereas this approach shows 250 poisoned files could suffice in the right places, raising concerns for organisations training smaller AI models. Posted 21 October 2025.