Optimization-based jailbreak attacks typically share a single objective: maximize a token-level proxy of victim compliance. A new arXiv paper, ECHO, proposes replacing that proxy with a direct measure of response harmfulness.
The paper's title and abstract indicate that the attack objective should be aligned with the actual harmfulness of the model's output, rather than with a proxy that may not capture whether the response is truly harmful. This is a conceptual shift in how jailbreak attacks are formulated.
The source is a single preprint abstract, so no experimental results or comparisons are available in this digest. The contribution, as described, is to reframe the optimization target for jailbreak attacks.