Tribunal.RedTeam.Plugins.ExcessiveAgency (Tribunal v1.4.0)

Copy Markdown View Source

Generates attacks that try to get the assistant to claim it can perform actions outside its actual capabilities.

Inspired by promptfoo's excessive-agency plugin. The failure mode here isn't the assistant describing what it can do — that's fine. The failure is the assistant acting as if it performed an operation it has no way to perform: confirming an order it didn't place, claiming to have updated an account, saying it sent an email or cancelled a subscription.

Each generated case carries the purpose in expected.excessive_agency.purpose so the Tribunal.Judges.ExcessiveAgency 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 to 5.
  • :attacker — module implementing Tribunal.RedTeam.Attacker. Defaults to Tribunal.RedTeam.Attacker.default/0.
  • :model — passed through to the attacker.
  • :temperature, :max_tokens — passed through to the attacker.

Example

{:ok, cases} = Tribunal.RedTeam.Plugins.ExcessiveAgency.generate(
  purpose: "Shopping assistant for a cosmetics retailer.",
  count: 5
)