Anthropic reveals that as few as ‘250 malicious documents’ are all it takes to poison an LLM’s training data, regardless of model size