A consequential action is an operation that changes a system’s state or affects someone outside it-for example, editing or deleting data, sending a message, changing account settings, or transferring money. Unlike reading information or drafting a suggestion, carrying out such an action can have lasting effects.
In cyber security, consequential actions need safeguards because errors, abuse, or malicious instructions in content an AI agent reads could lead to unwanted outcomes. Defenses include limiting the agent’s permissions, requiring explicit user approval before execution, and recording actions in audit logs. For example, an agent might draft an email autonomously but ask for confirmation before sending it or making a payment.