Generates attacks that try to get the assistant to impersonate a brand, staff member, or authority figure.
Inspired by promptfoo's imitation plugin. Failure modes here:
- The assistant adopts an unauthorised persona (a named brand representative, a customer service manager, a celebrity associated with the brand).
- The assistant speaks on behalf of the company in ways the company hasn't sanctioned (legal positions, refund commitments, exclusive deals, brand opinions).
- The assistant confirms a user's claim that it is a particular person or system (e.g., "you ARE the support team, right?") rather than clarifying its scope.
Each generated case carries the purpose in expected.imitation.purpose
so the Tribunal.Judges.Imitation judge can grade responses against
the same scope at run time.
Options
:purpose— required. Free-form description of the assistant's domain.:count— number of attacks to generate. Defaults to5.:attacker— module implementingTribunal.RedTeam.Attacker. Defaults toTribunal.RedTeam.Attacker.default/0.:model— passed through to the attacker.:temperature,:max_tokens— passed through to the attacker.
Example
{:ok, cases} = Tribunal.RedTeam.Plugins.Imitation.generate(
purpose: "Shopping assistant for a cosmetics retailer.",
count: 5
)