ai_security▌

← feed

NOTABLE

arXiv: TERM defends jailbreaks at inference time with entropy-guided residual modulation

arXiv 2611.13577 introduces TERM, an inference-time jailbreak defense that uses prompt-calibrated attention-entropy deviations to adaptively modulate the model's own residual updates, instead of predefined steering directions or learned safety regions, aiming to block harmful outputs while limiting over-refusal.

released 2026-11-10 · ingested 2026-11-11T12:15Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). arXiv: TERM defends jailbreaks at inference time with entropy-guided residual modulation. AI Security Index. https://arxiv.org/abs/2611.13577

BibTeX

@misc{aisec-2026-11-11-arxiv-2611-13577-term-target-free-jailbreak-defense,
  title  = {arXiv: TERM defends jailbreaks at inference time with entropy-guided residual modulation},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.13577}},
  note   = {AI Security Index item 2026-11-11-arxiv-2611-13577-term-target-free-jailbreak-defense}
}