Prompt Optimization and Evaluation for LLM Automated Red Teaming

Applications that use Large Language Models (LLMs) are becoming widespread, making the identification of system vulnerabilities increasingly important. Automated Red Teaming accelerates this effort by using an LLM to generate and execute attacks against target systems. Attack generators are evaluated using the Attack Success Rate (ASR) the sample mean calculated over the judgment of success for each attack. In this paper, we introduce a method for optimizing attack generator prompts that applies ASR to individual attacks. By repeating each attack multiple times against a randomly seeded target, we measure an attack's discoverability the expectation of the individual attack success. This approach reveals exploitable patterns that inform prompt optimization, ultimately enabling more robust evaluation and refinement of generators.

Don't Lie to Me: Avoiding Malicious Explanations with STEALTH

STEALTH is a method for using some AI-generated model, without suffering from malicious attacks (i.e. lying) or associated unfairness issues. After recursively bi-clustering the data, STEALTH system asks the AI model a limited number of queries about class labels. STEALTH asks so few queries (1 per data cluster) that malicious algorithms (a) cannot detect its operation, nor (b) know when to lie.