Security researcher Micah Lee combined personal criticism of major AI companies with a proposal to use open-weight models through other inference services. His experience with GLM 5.3 is a subjective assessment, not a formal comparison. Related Tinfoil documentation describes protecting data in use and executing code through secure execution environments and remote attestation. Limits remain, including trust in hardware vendors and side channels. Training opt-out settings, model openness, and an inference operator’s access controls should be evaluated separately.