Hackers attack AI servers today. Soon they will attack the model itself, and your agent will do exactly what the hacker wants with zero bugs in the code. Screenshot this. Today they go after the plumbing: SGLang, the serving engine under xAI, Cursor, Oracle and the big clouds, had three critical RCEs this year. The weights are untouched because the Python is easier. Next they go after the weights. Delta League just showed that 0.016% of a model's numbers decide how it behaves, no training needed. An attacker inside an inference server does not have to steal the model. He edits a few thousand weights in GPU memory and leaves. Your trading agent keeps running, slightly wrong, in his favor. Attestation proves what was loaded, not what is in memory an hour later. So stop trusting the model to protect you. NEAR AI is adding a turnstile inside the enclave that flags when anything unusual runs. OutLayer keeps the agent's wallet under a policy the model never touches: a poisoned model can suggest the trade, it cannot approve it.