AI Model Vulnerability - Just 250 poisoned documents can corrupt AI models with up to 13 billion parameters [1] Security Implications - The study highlights the need for new kinds of defenses against data poisoning attacks on AI models [1]

New Anthropic research: We found that just a few malicious documents can produce vulnerabilities in an AI model—regardless of the size of the model or its training data.This means that data-poisoning attacks might be more practical than previously believed. https://t.co/YMod3czB4X ...