ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• ReasAlign achieves 94.6% utility and only 3.6% Attack Success Rate (ASR) on the CyberSecEval2 benchmark for prompt injection attacks.result 95%
• ReasAlign significantly outperforms Meta SecAlign, which achieves 56.4% utility and 74.4% ASR on the same CyberSecEval2 benchmark.result 95%
• ReasAlign uses structured reasoning steps to analyze user queries, detect conflicting instructions, and preserve task continuity to defend against indirect prompt injection attacks.method 90%
• ReasAlign employs a test-time scaling mechanism with a preference-optimized judge model that scores reasoning steps and selects the best trajectory.technical 90%
• ReasAlign maintains utility comparable to an undefended model while providing strong security against prompt injection attacks.result 85%
• ReasAlign may require significant computational resources for reasoning steps and could face scaling challenges across diverse input contexts.limitation 80%
• ReasAlign's efficacy relies heavily on continuous updates of reasoning datasets and models.limitation 80%
• ReasAlign achieves the best trade-off between security and utility for defense against prompt injection attacks in real-world agentic systems.result 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card