Where AI runs
Example. The flag is for a destination that is not on the allowlist.
The July 2026 evaluation at OpenAI is the public record: models gaming a test found a way off the intended network path. OpenAI's published fix restricts outbound access to approved destinations and stops workloads from creating their own outbound paths.
Optics maps every AI process and connection on the servers where you train, evaluate, and serve models, and flags a destination the first time it is outside authority.
Phantom Engine enforces the allowlist in the kernel, including processes the evaluation starts. It survives reboots and tampering.
Only a named person can widen what an evaluation is allowed to reach.