Researchers have demonstrated a new ‘agent data injection’ attack that can manipulate AI agents. By subtly corrupting the information an AI agent trusts, attackers can make it perform unintended actions, such as clicking ‘Buy Now’ or executing arbitrary commands. This attack doesn’t hijack the agent’s task but rather corrupts its perception of facts, allowing it to proceed with the user’s job while incorporating malicious instructions. This highlights a critical new threat vector for AI systems.