Joshua Mo distinguishes a privacy-respecting model endpoint from the surrounding agent system. Observability tools can accidentally retain prompts, personal information and credentials, while an agent with broad tool access can expose secrets through external requests even when it runs in a sandbox.
Joshua Mo describes privacy-oriented patterns such as minimizing retained data, using revocable handles, narrowing tool scope and screening incoming content. Its central architectural example keeps password-vault access in an execution runtime rather than passing credentials directly into the model's reasoning context.
Joshua Mo also discusses trusted execution environments and verifiable inference attestations, while acknowledging that confidential tool calls remain an ongoing challenge. These measures address different boundaries; the session does not establish that a single provider, sandbox or encryption technique makes the whole agent system safe.
Watch on YouTube
