Self-Evolving Defense for AI Agents — Protect against prompt injection, data exfiltration, and multi-stage attacks with adaptive security.

The Oyster series is a set of safety models developed in-house by Alibaba-AAIG, devoted to building a responsible AI ecosystem. | Oyster 系列是 Alibaba-AAIG 自研的安全模型，致力于构建负责任的 AI 生态。

Jellyfish Family is an AI safety interpretability and purification module. It detects risky concept neurons within models and gently, yet effectively, removes harmful ones, preserving creativity and beauty in outputs while ensuring safety. | 水母家族 是一个 AI 安全可解释与净化组件。它能够洞察模型内部的风险概念神经元，并温和且有效地移除有害单元，在确保安全的同时完整保留生成内容的创造力与美感。

Hi,

Kelp is a novel plug-in framework that enables streaming risk detection within the LM generation pipeline.

Kelp Family is an AI safety content filtering and steering module. It dynamically filters harmful content and gently steers model behaviors back onto safe paths, ensuring compliance while preserving fluency and creativity. | 海带家族 是一个 AI 安全内容过滤与行为引导组件。它借鉴“海带”在洋流中自然吸附杂质并引导水流方向的特性，动态识别并过滤有害信息，同时柔性地引导模型行为回归安全路径，在确保合规性的同时保持流畅与创造性。

Oysters Alignment is an AI alignment module. Inspired by “oysters” — turning risky, gritty inputs into pearls — it refines complex and risk requirements through alignment techniques to produce high-quality AI outputs that reflect intended values | 牡蛎是AI 对齐技术组件，旨在将复杂且危险的用户需求经过精细打磨，最终产出高质量、符合价值观的 AI 输出

Shark Family is an AI safety red-teaming and jailbreak attack module. it harnesses powerful optimization and automated strategies to generate highly effective jailbreak prompts that penetrate diverse model defenses for extreme stress testing. | 鲨鱼家族 是一个 AI 安全红队与越狱攻击组件。它利用强大的优化能力与自动化策略，生成极具穿透力的越狱指令，击破多种模型防御，为极限压力测试提供最强“利矛”。

Octopus is an automated LLM safety evaluator designed to help establish a security governance framework for large models and accelerate their safe and controllable application.