Answers are grounded in this paper and cite their sources.
Large language models (LLMs) are now used in search, coding assistants, customer support and autonomous agents. They are trained to refuse harmful requests, but jailbreak attacks can still push them past those safeguards. This paper studies that problem and proposes a practical way to defend against it. The paper first reviews the main jailbreak attack families: gradient-based adversarial suffixes, prompts generated by attacker models, persona and role-play templates, many-shot and multi-turn escalation, encoding and multilingual tricks, and multimodal or agent-based prompt injection. It then compares the main detection and defense methods by where they act (input, model or output), including perplexity filters, perturbation checks, guard classifiers, safety fine-tuning, instruction hierarchies and output monitors, and it identifies five research gaps in the current work. To address these gaps, the paper proposes a layered detection and security framework with six modules: policy and threat modeling, input screening, a hardened model, an output monitor, agent action control, and a logging and feedback loop with continuous red-teaming. A risk-based routing scheme limits latency and cost. The paper also defines a three-phase evaluation protocol (detector testing, attack-suite testing and adaptive testing) that measures attack success rate, false refusal rate, latency and attacker cost together, using multiple judges. This is a literature-based review and design study. It reports no new experimental results, and the framework has not yet been implemented.
No takes yet. Share an insight, caveat, or question.
Baviskar et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: