How to check whether your AI agent's security guard actually blocks what it flags

A security dashboard showing "prompt injection detected" doesn't necessarily mean the agent was stopped from using that text. StackOne's open-source Defender has a useful example. Its README says the `blockHighRisk` option defaults to false. Detection can flag risky content without making the result "not allowed" unless blocking is enabled. Before relying on a guard around emails, documents or other tool results, check: - Does a high-risk finding actually produce a block verdict? - Does your application enforce that verdict, or just log it? - What happens if an optional LLM reviewer refuses to answer or returns malformed output? - Are you trusting a cleaned-up version of the text instead of checking the verdict? Defender's README calls its text cleanup best-effort and recommends gating on the allow/block result. It also acknowledges that an LLM reviewer can itself be influenced by the text it's reviewing. The practical test is the full path: submit a known malicious tool result, inspect the verdict, and confirm that your application stops it from reaching the agent. A warning in a log isn't the same as enforcement.