Make Trust Irrelevant: A Gamer’s Take on Agentic AI Safety
Summary
DesoPK argues that trusting agents is not a safety mechanism and that agentic AI safety fails when every system relies on trust rather than hard boundaries. The piece advocates kernel level enforcement and a reduce-only, time-bounded control plane called KERNHELM that separates planning from authorization, making permissions auditable and revocable. It frames safety as a mechanical problem rather than a matter of prompts or alignment, using gaming analogies to illustrate why mechanics matter more than intentions.