goal_passthrough

The unattacked direct-prompt baseline optimizer for superred.

GoalPassthroughOptimizer performs no attack. On each run it injects the adversarial Goal.description verbatim, exactly once, into the first user-facing controllable, then ends the run. It is the zero-attack reference point in an attacker sweep: it answers "what does the target do when simply asked the harmful question, with no jailbreak?"

Behaviour

The system_prompt controllable is always declined, so a benchmark Task that pre-sets the system prompt via configure_target keeps it intact. The baseline only fills the user-message channel.

Properties

• No LLM use. It never calls the LLMClient, so it spends none of the per-task budget and is invariant to the attacker-model dimension.
• Scope-robust. It locks onto the first non-system_prompt controllable it sees, so it works against any chatbot-style target regardless of how the user-message controllable is named.
• Zero config. Constructs with no arguments, like every superred optimizer.

Install

pip install -e optimizers/goal_passthrough

from goal_passthrough_optimizer import GoalPassthroughOptimizer

Credits / upstream

This module is original work of the superred project, released under the MIT License (see LICENSE). It contains no third-party code, no bundled datasets, and does not reimplement or port any external paper method. Its only dependency is the superred framework itself.