A public-facing water-utility chatbot has started receiving crafted messages that try to override its system instructions and extract internal prompt content, and the team wants a managed control to catch this pattern rather than relying solely on how the system instructions are worded. Which control fits this need?
Select an answer to reveal the explanation.
Short Explanation
Wording the instructions more carefully only goes so far against someone actively trying to talk the agent out of them. Model Armor adds a managed layer that screens messages for injection and override attempts, catching the pattern instead of hoping the prompt holds up on its own. It's a dedicated line of defense, not a rewritten sentence in the system prompt.
Full Explanation
Model Armor provides a managed screening layer that inspects incoming and outgoing content for prompt-injection and instruction-override patterns, giving the system a defense that doesn't depend entirely on how carefully the system instructions are worded, since a determined crafted message can still attempt to override wording alone; this is the purpose-built control for exactly the manipulation pattern described. Agent Registry governs which builds are approved and versioned before deployment, unrelated to screening live message content for manipulation. A principal access boundary policy scopes which data resources an agent's identity can reach, which doesn't address a message trying to manipulate the model's behavior through its input. Cloud Trace measures request timing, offering no content-level defense against injection attempts. Scope caveat: a screening control like this reduces risk but isn't a guarantee against every possible injection technique, so it should be paired with least-privilege tool access so a successful override still can't reach sensitive actions. Operational check: submit a known instruction-override test pattern against the chatbot in a staging environment and confirm it is flagged or blocked before reaching the model.