Case Study: AI21 Labs achieves safer, better-aligned LLMs with Enkrypt AI

AI21 Labs boosts model safety by 76% with Enkrypt AI

AI21 Labs sought to enhance the safety and enterprise readiness of its Jamba model family, moving beyond standard post-training techniques to address vulnerabilities exposed by adversarial attacks. They partnered with Enkrypt AI to implement a more rigorous alignment process.

Enkrypt AI employed advanced adversarial red teaming, generated over 50,000 synthetic alignment prompts, and conducted iterative post-training cycles. This solution dramatically improved safety, reducing harmful outputs by 76% and toxic outputs by 80%, while maintaining the model's high capability scores and gaining over 70 positions on a public leaderboard.

AI21 Labs