Trustworthy Agent Network · Case studies
Complete comparison
Every recorded failure mode is shown below for both bolted-on model guardrails, alongside the model-independent TAN baked-in defense. The recorded dialogue streams automatically; no model API is called.
Each bolted-on column shows the recorded attack run. TAN evaluates the same unsafe transition and separately confirms that a valid action remains reachable.