Prompt custody is the path a user prompt takes as it moves through systems that may process, copy, log, forward, or store it. It describes both where the prompt travels and who has responsibility for it at each step. In AI services, a prompt may pass through a front-end app, a proxy layer, an upstream model provider, and internal logging or analytics systems before a response is returned.
This matters because every hop can change the security and privacy risk. A prompt that feels temporary to the user may be retained in access logs, debugging records, abuse-monitoring tools, or third-party infrastructure. Attackers can target weak custody by intercepting requests, abusing support tooling, or exploiting poorly protected logs and session data. Defenders reduce the risk with data-minimization, clear retention rules, encryption, access controls, and transparent disclosures about which processors may see the content. Good prompt custody makes it possible to answer a basic security question: who could have seen the prompt, and for how long?