SpecGuard monitors multi-turn LLM jailbreaks with temporal logic
Original title: SpecGuard: Detecting Multi-Turn LLM Jailbreaks via Specification-Based Runtime Monitoring
A training-free monitor compiles Linear Temporal Logic rules into automata and emits replayable violation witnesses.
SpecGuard is a training-free framework for detecting multi-turn LLM jailbreaks by treating dialogue safety as a trace property. It maps observations to evidence-backed facts under a fixed atomic-proposition contract and checks forbidden patterns written as Linear Temporal Logic specifications. Compiled symbolic automata support buffered runtime checking and retrospective auditing, and accepting runs produce replayable witnesses. On multi-turn safety benchmarks, the authors say it outperforms representative open-source guard models with controlled false-positive rates, without learned detector parameters or access to the protected model.
Unconfirmed: Benchmark names and the size of the reported gains are not stated in the abstract.
- Violations may appear only across user intent and assistant behavior over several turns.
- Witnesses name the triggered specification, supporting facts, and relevant turns.
- The monitor does not require learned detector parameters or model access.
From the original article
1 min read at arxiv.orgMulti-turn jailbreaks turn dialogue safety into a trace property: a violation may become identifiable only when user intent and assistant behavior are considered across multiple turns. Existing guard models largely classify individual messages or exchanges and return risk scores or labels, without executable safety semantics or a formally checkable decision trail.
By Junle Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.