SecReEvalBench: A real-world scenario-based security resilience benchmark for large language models.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Natural Language Processing

Background:

• Current large language model (LLM) security evaluations often overlook multi-turn interactions.
• Vulnerabilities to prompt-chain attacks, which exploit context retention and sequencing, are underestimated in single-turn or fixed-domain settings.
• Sequence-aware adversaries pose a significant threat, necessitating advanced evaluation methods.

Purpose of the Study:

• To introduce SecReEvalBench, a novel benchmark for assessing LLM security resilience against prompt-chain attacks.
• To develop sequence-sensitive metrics for multi-turn LLM safety evaluations.
• To provide a standardized and reproducible dataset for advancing LLM security research.

Main Methods:

• Developed SecReEvalBench with six distinct attack sequences and four tailored multi-turn security metrics.
• Created a comprehensive dataset spanning seven security domains, sixteen attack techniques, and four maliciousness levels.
• Employed dual-LLM adjudication for intent labels and combined an unsafe-content detector with a refusal classifier for nuanced measurements.

Main Results:

• Sequence-aware metrics exposed vulnerabilities missed by single-turn tests in both open-weight and proprietary LLMs.
• Identified issues such as delayed or absent refusals during attack escalation and information leakage through intermediate reasoning.
• Demonstrated that current safety measures can be bypassed by sophisticated, multi-turn adversarial prompts.

Conclusions:

• SecReEvalBench effectively standardizes multi-turn LLM safety assessment.
• The benchmark provides a reproducible foundation for developing defenses against domain- and sequence-aware attacks.
• Publicly available dataset facilitates further research into robust LLM security.