The work models how API secrets introduced into an LLM agent can spread through prompts, histories, logs, memory stores, generated code, and errors, then be abused through or excessive agency. It proposes keeping authentication outside the model: the model chooses a connector, while a trusted request boundary adds the credential.
Two controlled black-box experiments evaluated a production implementation across 16 probes in seven control domains. An authenticated GitHub request succeeded without exposing the credential through the tested environment, headers, filesystem locations, echo services, or unrelated origins, and cloud metadata endpoints were unreachable. A misconfigured connector also showed that secure custody does not guarantee working authentication. The result supports vault mediation as one security layer, not a substitute for least privilege, action authorization, approvals, redaction, and rotation.