Preserved thinking is a safety control that keeps a model’s internal reasoning context protected from user manipulation through an API. In practice, it means callers should not be able to edit, erase, or directly extract the model’s private working state while a request is running. The goal is to preserve the integrity of the model’s hidden deliberation so the output reflects the system’s intended behavior, not attacker-supplied tampering.
This matters in cyber security because internal context can become a target for prompt injection, data exfiltration, and workflow abuse. An attacker may try to coerce an agent into revealing hidden prompts, reasoning traces, or tool plans, or to alter that context so later steps behave unsafely. Preserved thinking is a defense-in-depth measure: it narrows one attack path, but it does not replace sandboxing, least privilege, output validation, or human review for sensitive actions.