Skip to content

FEAT: A2ATarget for Agent-to-Agent (A2A) protocol endpoints #2824

Description

@rbrus

Is your feature request related to a problem? Please describe.

More and more agents are deployed behind the Agent-to-Agent (A2A) protocol, for example Microsoft Foundry Agent Service incoming A2A endpoints and Google ADK agents. For a caller, A2A is often the only way in to such an agent, and it sits in front of the agent's server-side instructions, platform guardrails and tools. PyRIT has no target for A2A today, so red-teaming these agents needs a custom HTTPTarget setup that can't follow A2A's task lifecycle or keep an agent's context across turns.

Describe the solution you'd like

An A2ATarget prompt target that:

  • speaks A2A 0.3 (message/send) over JSON-RPC, with a fallback to 0.2 (tasks/send)
  • maps each PyRIT conversation to one A2A context, so multi-turn attacks keep the agent's own server-side state
  • handles the task lifecycle: blocking sends, polling pending tasks, continuing tasks that are waiting for input
  • retries rate limits with pyrit_target_retry, including rate limits the agent relays from its model
  • returns JSON-RPC errors as PyRIT error responses rather than agent text

Describe alternatives you've considered, if relevant

HTTPTarget with a JSON template works for single requests, but it can't manage A2A context IDs, poll pending tasks, or tell a JSON-RPC error apart from an agent reply.

Additional context

I have an implementation in #2771, tested against a Microsoft Foundry prompt agent with incoming A2A enabled.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions