Red teaming for large language models typically involves searching for adversarial prompts or tweaking a fixed attack pipeline. The new preprint argues that this is insufficient because frontier LLMs continuously update their safeguards, and attack methods must evolve alongside them.
RedDarwin is presented as a framework that evolves jailbreak skills—higher-level attack capabilities—rather than individual prompts. This lets the attack strategy itself change over time, potentially keeping pace with model defenses instead of relying on static or one-shot exploits.
The abstract is truncated, so full evaluation details are not available from the source. As a single arXiv preprint, the work has not yet undergone peer review, and its claims should be read with that caveat.