10 Things to Know About AI Agents and Their Safety Risks
Analytics Insight, Friday, September 11th, 2026
A photo explainer on how AI agents work and the safety risks they bring, from excessive permissions to prompt injection.
AI agents are moving beyond answering questions to planning tasks, using tools and taking actions with limited human oversight, which creates new safety challenges. Unlike chatbots, agents can send messages, access information and update systems on a user's behalf.
Granting more permissions than necessary increases potential damage, and attackers can plant malicious instructions in webpages or documents via prompt injection.
Agents handling emails and customer records need strong access controls, can misread instructions or pick the wrong tool, and each connected tool opens another attack path through an organization's systems.