Model Identity Protection

Model Identity Protection is WitnessAI's model identity assurance guardrail. It gives you two controls: it can keep instructing a private model about its own identity, which makes it harder for user prompts to change that identity, and it checks the model's responses to confirm they stay consistent with it.

When a response matches one of your rules, the guardrail can Allow, Warn, Block, or Route it, with a customizable message.

WitnessAI Policies enable organizations to control, restrict, and protect the use of AI Models and Applications. Based on user activities, policies can block prompts, route usage to preferred AI Models or Applications, warn users, and maintain compliance with security and usage policies.

How it works

You shape this guardrail with rules, the same way as the Behavioral Activity guardrail. Each rule has two parts:

A condition, shown as Model Response: one or more intentions the rule looks for. Each intention has an editable Definition.
A THEN outcome: a severity (High, Medium or Low) and an action (Allow, Warn, Block or Route).

The last rule, For all other:, sets the outcome for responses that none of your rules match: Allow or Block.

The optional Model setting is the first of the two controls. It continually supplies an instruction to a private model to reinforce its identity. The rules are the second control, because they check what the model sends back.

Using Model Identity Protection step-by-step

Starting within an existing or new Policy:

Click the Guardrails tab.
Select Model Identity Protection from the list of guardrails.
Click the toggle to enable the guardrail.
Optional: under Model (optional), select Set context for model, choose a model and enter the instruction that defines its identity, such as who the model is and how it should describe itself. Only private models are listed.
Optional: describe the policy in the Purpose field. With no rules yet, click Create it for Me to have WitnessAI draft rules from the Purpose. Once you have rules, click Refine All Rules to improve them. Review the proposal and click Keep, Undo or Try Again.
To add a rule by hand, click Manually Add Rule (it reads Add Rule once a rule exists). Under Model Response, choose an intention and review its Definition. Click Add Intention to add more intentions to the same rule.
Under THEN, set the severity and the action. Enter the message to show the user, or choose the model for Route.
Set the For all other: rule to Allow or Block.
Save the policy.

Note: When a rule's action is Block or Warn, the console shows a Response buffering notice. Responses are held until the guardrail finishes evaluating them, which can affect the user experience. See Response Buffering.

The Model Identity Protection tab enabled, showing the Purpose field, a collapsed rule and the Refine All Rules control.

Model Identity Protection How it works Using Model Identity Protection step-by-step