SecureGARD inserted an external semantic policy check before agent tools ran

The approach records semantic state and checks candidate tool actions before effects are applied. Its decision is not simply whether a response looks suspicious; it tests whether the requested transition fits a known resource and authority model. The evaluation discusses both attacks blocked and legitimate task completion costs.

The result is a paper evaluation, not an independently audited production guard. Unknown domain state, incomplete ontologies and benign task interruptions limit any claim of fail-safe prevention.

The proposed guard evaluates the intended state transition at the action boundary, outside the language model's own explanation. Its ontology encodes permitted relationships and is updated offline, then frozen for the runtime evaluation. The authors report zero attack successes across their selected tasks and backbones, a measured sample result rather than evidence that every novel tool or incomplete ontology is covered. The crucial implementation question is what happens when the semantic state is missing or wrong, and whether benign work remains possible.