Precise words. Practical meaning.

A model given tools, a goal and the authority to act in a loop without a human approving each step, deciding for itself when it is finished.

A cryptographic statement, signed by hardware, describing exactly what code is running in an enclave. It lets a remote party verify an environment they do not control.

An append-only record of what a system did, on what input, under whose authority, and when. For AI systems this includes the prompt, the model, the response and the cost.

Running computation inside a hardware-enforced enclave so that even the operator of the machine cannot read the data being processed. Attestation lets a third party verify which code is running inside it.

Network traffic leaving a system boundary. In a sandboxed agent context, controlling egress means deciding which outbound destinations the agent may reach at all.

An EU AI Act classification covering systems used in areas such as employment, credit scoring, education, critical infrastructure and law enforcement. It triggers obligations on risk management, data governance, logging, human oversight and conformity assessment.

A design where a person reviews or approves an automated system's output before it takes effect. Named explicitly in EU AI Act obligations around human oversight for high-risk systems.

AI systems where the operator retains control over where the model runs, where the data goes, and who can inspect either. Often used for national or sector-level infrastructure, but it applies to any single organisation.