What AI says about GraySwanAI

GraySwanAI's Circuit Breakers project develops Circuit Breaking, a safeguard that directly alters model representations to prevent harmful content generation. The approach offers a family of methods as an alternative to refusals and adversarial training, aiming to protect LLMs and multimodal models from unseen adversarial attacks without sacrificing performance. The work, including a research paper and accompanying code and models, targets improved alignment and robustness of AI systems.