A safety practice for testing risky AI inside a controlled space.
AI Evaluation Containment is like a quiz room with no phones. The model sits down. The Wi-Fi stays off. Nice try, robot.
Teams use it before a risky model goes public. Secrets stay in. Off-limits moves stay blocked. The model cannot grade itself.
Third-party AI evaluation
Containment keeps outside tests clean and away from model tricks.
AI sandbox
An AI sandbox gives the test safe tools and network limits.
Frontier Model Access Control
Access control decides who can reach the frontier model under test.
AI Red Team
The riskier the red-team test, the stronger the walls need to be.