Compliance guards fail a crossed-rule audit

A new audit of AI compliance guards finds that deleting, swapping, or permuting the governing rule left every tested detector’s accuracy unchanged, exposing what the authors call “rule blindness.” The crossed-rule benchmark tested four deployed guard models, an 8B judge, and 13 benchmarks; its inexpensive Internal Compliance Score improved mechanically verified ranking until an adaptive white-box attack removed the gain. Teams should treat a guard’s citation of policy as evidence to test, not proof that the policy changed its decision.

That is the security counterpart to the site’s Astra release-gate analysis: a control is only real when an adversarial or counterfactual test can make it fail visibly.