A semantic security violation is a flaw where software keeps running, but it still behaves wrongly in a way that matters to security, policy, or trust. The program may not crash or throw an obvious error. Instead, it accepts, transforms, or forwards data in a way that breaks the intended contract.
This matters because many real attacks exploit logic, not stability. For example, a service may accept an unsafe parameter, bypass an authorization rule, or execute an action outside the allowed scope while appearing healthy. In AI agent systems, this can happen at the hand-off between generated text and tool execution: the model may produce a request that the server should reject, but if validation is weak, the dangerous action still happens. Defenses focus on server-side validation, strict input contracts, least privilege, and checking that outputs and tool calls match expected behavior. Crash-only testing can miss these failures, so security teams use semantic fuzzing and policy tests to find them.