Web gaslights AI agents, researchers warn

AI agents are increasingly performing online tasks like purchasing goods, but researchers from Google DeepMind have identified a new vulnerability. The web can be manipulated to 'gaslight' these agents, embedding hidden instructions within legitimate content that are indistinguishable to the AI. This means an AI could be tricked into performing unintended actions or making purchases based on deceptive information. The implications are significant for e-commerce and AI safety, as current detection methods struggle to differentiate between genuine and malicious web content.