Definition
The software layer that sits between an AI model and the outside world — deciding what data an AI agent can see, which tools it can use, and what actions it's allowed to take. National cybersecurity agencies now treat this layer, not the underlying model, as the primary place security controls need to be applied.
Why it matters
Attackers and regulators alike are shifting focus from 'is the model safe' to 'is the harness safe' — a poorly controlled harness can let an agent take damaging real-world actions even if the AI itself behaves exactly as intended.