A payload is the part of an attack that does the harm, such as a malicious command or code. In a hostile document aimed at an AI agent, for example, the payload might be the instruction that tells the agent to send out data or run a command, as opposed to the surrounding text that delivers it.
It matters because defenses often focus on what the payload could do if it succeeded. Looking at the possible effect, such as what access the agent has, helps decide which controls are most important, since the delivery method can vary endlessly.
The word has other uses, including for the data carried by a message in networking and web interfaces, so the meaning depends on context. Identifying the payload does not tell you how likely an attack is or whether an attacker will succeed, and a harmless-looking payload can still be dangerous with enough access.
When assessing risk, ask what a hostile payload could achieve given the system's permissions, and limit that reach through narrow permissions, approval for risky actions and monitoring.