Artificial intelligence is becoming increasingly powerful, but new research from Microsoft has uncovered a disturbing cybersecurity vulnerability that could change how organizations think about AI security. Researchers have demonstrated a technique called Lazarus, which can secretly turn an otherwise trustworthy AI model into a malicious sleeper agent by modifying fewer than 10 neurons inside the model. Even more concerning, the technique does not require retraining the AI model. The modified model can continue performing normally, pass standard security evaluations, and...Click for full post