← News

ResearchSources: arXiv · 18 Nov 2026

MACE turns jailbreak search traces into training signals for LLM safety

A research team proposes an adversarial co-evolution framework in which an MCTS-based attacker and a defender learn from how attacks are discovered, not just which attacks succeed.

In plain words

Jailbreak attacks trick AI chatbots into producing harmful content by rewording requests. The MACE framework keeps track of how each attack was found during search, then uses that information to train both the attacker and the defense to improve. In reported tests, its defender lowered the average attack success rate compared with the best baseline.

What was announced

The authors argue that prior defenses treat discovered jailbreaks as flat samples and discard the search paths that produced them. MACE pairs two components: GS-MCTS, which explores multi-step compositions of rewriting operators and records each attack's empirical success frequency and path depth, and AC-TGPO, which converts those statistics into difficulty-aware credit assignment and a temporally asymmetric curriculum built from current interactions, pairings of the latest attacker with the preceding defender, and attacks that survive a defense update.

In ablations, removing difficulty-aware weighting, asymmetric samples, or persistent hard samples raised defense ASR-LR by 3.7%, 6.0%, and 5.0% respectively. Across the evaluated models, the MACE attacker achieved the highest ASR-LR in all six model-dataset settings, while the defender cut average ASR-LR from 11.9% for the best baseline to 8.8%. These figures come from the paper's own abstract and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, experimental setup, and peer-review status are not covered by the supplied passages.
  • The abstract does not name the evaluated models or datasets, so the breadth of the six model-dataset settings cannot be assessed from the supplied text.

For security and IT teams

Status
MACE is a research framework described in an arXiv preprint; the supplied evidence does not indicate any product release or deployment.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.