A 311 chatbot keeps receiving messages that try to convince it to ignore its guardrails and reveal internal configuration details, and the team wants a filtering layer specifically aimed at catching this kind of manipulation. What should they add?
Select an answer to reveal the explanation.
Short Explanation
A jailbreak attempt is a message trying to talk its way past the rules, so you need something reading the message itself, not counting how many the sender sent. Model Armor is that reader, sitting in the conversation path looking for manipulation before it lands. A rate limit or an identity policy might slow down or restrict an attacker, but neither one actually reads what they're saying.
Full Explanation
Model Armor is built to inspect conversational content for manipulation patterns, including prompt injection and jailbreak attempts, which directly addresses messages crafted to talk an agent into ignoring its own guardrails. An Agent Gateway rate limit constrains how fast or how often a sender can send messages, which might slow a brute-force attempt but does nothing to recognize or block a single, well-crafted manipulative message that arrives at a normal pace. An Agent Identity policy governs which principal or service account the agent's underlying processes are permitted to assume, which is a resource-access control; it has no visibility into the semantic content of a conversation and wouldn't stop a manipulation attempt that never touches a restricted resource. A Sensitive Data Protection scan over stored conversation history operates after the fact on retained data, looking for sensitive content patterns, not for real-time manipulation attempts in an active conversation, so it would catch this problem far too late, if at all, since manipulation attempts aren't necessarily sensitive-data patterns. A scope caveat: content screening reduces but doesn't eliminate manipulation risk, since novel phrasing can sometimes evade a filter, so it should be paired with monitoring for unusual downstream behavior. As an operational check, the team can run a known jailbreak phrase through the pipeline and confirm it is flagged or blocked before the model responds to it.