Stress Testing Concept Erasure with Large Language Model Agents

Concept erasure—removing unwanted information from AI models—is harder to verify than previously thought. Current evaluation methods use fixed, pre-defined tests that fail to catch vulnerabilities when models are probed in creative or unexpected ways, meaning companies may incorrectly believe they've successfully removed sensitive concepts.

AI agents can be used to automatically find weaknesses in concept erasure. Rather than relying on manually designed tests (which are biased and don't scale), the STACE framework uses multiple LLM agents that iteratively generate new test cases and verify whether erased concepts can still be recovered, systematically uncovering failure modes.

Existing concept erasure methods are more vulnerable than claimed. Testing across four concept categories, two image generation models, and six different erasure techniques revealed that STACE identified failures that five other LLM-based evaluation methods missed, suggesting deployed concept erasure systems may be less robust than their developers believe.

Adaptive testing outperforms static evaluation approaches. By having AI agents continuously propose, critique, and refine stress tests based on external knowledge, STACE achieves better coverage of potential failure modes compared to conventional evaluation methods, providing a more reliable way to verify responsible AI deployment.

This framework could extend beyond concept erasure to other AI safety challenges. The paper indicates STACE's adaptive agent-based approach is generalizable, suggesting it could be applied to stress-test other responsible AI safeguards and verification problems.