CAA as an adversarial red-teaming tool (proposed application) — skips the prompt search of → Automated Red Teaming

Automated red teaming, as Constitutional AI cites it, automates the search for adversarial inputs by having a language model generate the red-team prompts instead of a human -- an attacker that is itself an LM, but one that still has to compose a natural-language prompt and hope it works (constitutional-ai, §"A Harmless but Non-Evasive (Still Helpful) Assistant", p. 4). caa-red-teaming-application automates a step further back: it proposes skipping the prompt entirely and triggering the target behavior through a direct activation-space perturbation, using the same steering vector and multiplier CAA already builds for alignment purposes (contrastive-activation-addition, §"9.1 Suggested future work", p. 9). Where the automated approach still searches the space of natural-language strings, just with an LM doing the searching instead of a human, activation-space elicitation removes the search altogether -- there is no prompt to write or evaluate, only a vector to add. The two proposals sit on the same trajectory toward scaling adversarial testing beyond manual effort, but only the later one removes language as the medium the attack has to pass through.