Automated Jailbreaks of Large Language Models
Ensuring the security of large language models (LLMs) is no small task. One of the most challenging and resource-intensive things in LLM red teaming is identifying jailbreaks — prompts that bypass a model's safety mechanisms to produce harmful or prohibited responses. Finding these prompts is essential for testing the resilience of LLM-based systems, yet it often requires deep domain expertise and extensive manual effort.
But what if AI could test AI? Specifically, could companies automate the search for jailbreaks by having one AI system attack another in a controlled, repeatable way?
In this white paper, we evaluate two existing frameworks — tree-of-attack with pruning (TAP) and Declarative Self‑Improving Python (DSPy) — that approach this problem from different angles. We tested both frameworks in practical experiments to measure their effectiveness and identify their limitations and analyzed the results.
Download the white paper to find out more.
