Holdline: a neutral benchmark that catches whether AI agent guardrails actually work
A new MIT benchmark scores any AI agent write-guard on catch rate, false-block rate, and class-balanced Cohen's kappa, plus an injection-attack class that tries to talk the guard out of its verdict. It gives the DeepSeek Harness's 20+ guard plugins a shared way to prove they work, not just claim it. Re-runnable on any local model.