🤖 AI Summary
This study addresses the vulnerability of large language models to prompt injection attacks when sensitive information is embedded in system prompts, which can lead to unintended secret disclosure. The authors propose an adaptive adversarial attack framework that dynamically evolves attack strategies over more than 20,000 red-team evaluations to systematically assess nine state-of-the-art defense mechanisms. Experimental results demonstrate that all defenses relying solely on the model’s intrinsic safeguards fail to prevent information leakage, whereas only application-layer output filtering combined with hard-coded rules achieves zero leakage. This work provides the first large-scale empirical evidence—through adaptive adversarial testing—that security boundaries must be enforced by application code rather than by the model itself.
📝 Abstract
LLM-powered applications routinely embed secrets in system prompts, yet models can be tricked into revealing them. We built an adaptive attacker that evolves its strategies over hundreds of rounds and tested it against nine defense configurations across more than 20,000 attacks. Every defense that relied on the model to protect itself eventually broke. The only defense that held was output filtering, which checks the model's responses via hardcoded rules in separate application code before they reach the user, achieving zero leaks across 15,000 attacks. These results demonstrate that security boundaries must be enforced in application code, not by the model being attacked. Until such defenses are verified by tools like Swept AI, AI systems handling sensitive operations should be restricted to internal, trusted personnel.