WhitzardAgent is the open-source and research ecosystem of Whitzard.

We build open infrastructure for developing, evaluating, monitoring, and securing autonomous AI systems — spanning agent frameworks, runtime security, safety models, evaluation infrastructure, agent trajectory tooling, cybersecurity, and frontier AI safety research. Supported by Fudan University, NUWA Lab, Shanghai Innovation Institute and Shanghai Pudong Research Institute of Cryptology.

WhitzardAgent 是 Whitzard 白泽 的开源与研究生态，得到复旦大学、女娲实验室、上海创智学院、上海浦东密码研究院大力支持。

我们致力于构建面向自主智能系统的开放技术基础设施，覆盖 智能体框架、运行时安全、安全模型、安全评测、智能体轨迹基础设施、网络安全以及前沿 AI 安全研究，推动更强大的 AI 系统实现安全、可控的自主能力。

Security controls, monitoring, and containment for agents in action. 面向智能体实际运行过程的安全控制、监测与风险阻断。

Zero-Trust Security Foundation for AI Agents

A modular runtime security framework that can inspect and intervene across the agent execution lifecycle, including LLM calls, model outputs, tool invocations, and post-execution auditing.

面向 AI 智能体的零信任运行时安全框架，可在模型调用、模型输出、工具调用及执行后审计等多个阶段进行安全检查和干预。

Local-First Security for AI Agents

A local-first security layer for coding and general-purpose agents, combining rule-based checks, lightweight models, LLM-based guards, tool-call inspection, configuration scanning, and fail-closed protection.

面向编码智能体及通用智能体的本地优先安全层，融合规则、小模型、大模型安全检测、工具调用检查、配置扫描与 fail-closed 防护机制。

Lightweight models for reasoning safety, behavioral alignment, trust, and intent analysis. 面向推理安全、行为对齐、信任与意图分析的轻量化安全模型。

Think Twice Before You Act.

A plug-and-play safety model that corrects unsafe reasoning before it propagates into risky agent behavior.

通过在智能体行为执行前修正不安全推理，将安全防护前移到“思考—行动”之间。

GitHub · Hugging Face · Paper

Train in the MirrorWorld, Act in the Wild.

A simulation-to-real reasoning-correction framework for safer computer-use agents, transferring safety knowledge learned from simulated trajectories to real GUI environments.

面向 Computer-Use Agent 的仿真到真实推理校正框架，通过模拟环境生成安全经验，并迁移至真实 GUI 智能体。

GitHub · Hugging Face · Paper

Safety Detection over Reasoning Traces

A lightweight moderation model specialized for identifying hidden safety risks in intermediate reasoning traces of large reasoning models.

针对大推理模型中间推理过程的轻量化安全检测模型，用于识别隐藏在 reasoning traces 中的潜在风险。

GitHub · Hugging Face · Paper

Specialized models for studying agent intent, deception, manipulation, and human–AI trust in multi-turn interactions.

面向多轮人机交互中的 智能体意图、欺骗、操控与人机信任 的专项安全模型。

IntentNet · TrustNet

Infrastructure for reproducible agent safety evaluation and frontier-risk measurement. 面向智能体安全与前沿风险的可复现评测基础设施。

A unified evaluation runtime for safety benchmarks, agent evaluations, risk testing, scoring, and experiment management.

统一的 AI 安全评测运行框架，为安全基准、智能体评测、风险测试、评分与实验管理提供公共基础设施。

Official benchmark adapters and evaluation suites for WhitzardEval.

WhitzardEval 的官方 Benchmark 适配与安全评测集合。

Measurement infrastructure for studying the cybersecurity capabilities and behaviors of LLM-powered autonomous agents.

用于研究大模型智能体网络安全能力与行为特征的评测基础设施。

Frameworks, representations, simulators, and developer tools for building and studying agents. 用于构建、分析和研究智能体的框架、表示方法、模拟器与开发工具。

An Agent-Native Research Framework

A research-oriented framework for rapidly implementing agent methods, running benchmarks, and inspecting long-horizon trajectories through a unified execution kernel.

面向智能体研究的 Agent-Native 框架，在统一执行内核上支持智能体算法实现、Benchmark 运行以及长程轨迹分析。

The LLVM for Agent Trajectories

Compiler infrastructure that transforms heterogeneous agent traces into a canonical intermediate representation for training, evaluation, replay, verification, transformation, and observability.

面向智能体轨迹的编译基础设施，将不同 Agent Framework 和 Harness 产生的异构轨迹统一转换为中间表示，用于训练、评测、重放、验证、转换与可观测性分析。

LLM-based GUI Simulator

A controllable desktop-like simulation environment for synthesizing and evaluating GUI-agent interaction trajectories.

基于大模型的 GUI 模拟环境，用于生成和评测桌面智能体交互轨迹。

Open pipelines, datasets, and agent infrastructure for building and evaluating AI-powered cybersecurity systems. 面向 AI 网络安全模型与智能体训练、评测的开放数据、工具链与基础设施。

A scalable corpus-mining and quality-filtering pipeline for extracting high-quality cybersecurity training data from large-scale web corpora.

面向大规模互联网语料的网络安全数据挖掘与质量过滤流水线，用于构建高质量网络安全训练数据。

Open data foundations for training cybersecurity models and agents.

面向网络安全基础模型和智能体训练的开放数据基础。

• CyberSecurity-100B — Large-scale bilingual cybersecurity continual-pretraining corpus 大规模中英双语网络安全持续预训练语料

• CyberSecurity-1M — Curated cybersecurity knowledge dataset 百万规模高质量网络安全知识数据集

• CyberRepo-10K — Real-world vulnerability repositories and auditing tasks 面向真实软件漏洞与代码审计任务的代码仓库数据集

Explore more on our Hugging Face organization.

更多模型与数据集请访问 WhitzardAgent Hugging Face。

Our open-source work is connected to broader research on frontier AI risk, agent safety, systems security, cybersecurity, privacy, and trustworthy AI.

我们的开源项目同时服务于更广泛的前沿研究，包括 前沿 AI 风险、智能体安全、系统安全、网络安全、隐私与可信 AI。

Research on frontier AI safety is advanced through NUWA Lab, Whitzard's frontier AI safety research initiative.

前沿 AI 安全研究主要通过 NUWA Lab 女娲实验室 持续推进。

• Whitzard: https://whitzard.tech/
• Open Ecosystem: https://whitzard.tech/en/open-ecosystem/
• Hugging Face: https://huggingface.co/WhitzardAgent
• NUWA Lab: https://whitzard.tech/en/research/

We welcome researchers and developers interested in AI agents, AI safety, security, evaluation, and open infrastructure to contribute through issues, pull requests, and research collaborations.

欢迎关注 AI 智能体、AI 安全、安全评测与开放基础设施 的研究者和开发者通过 Issue、Pull Request 与研究合作参与我们的开源社区。

Research & Collaboration: contact@whitzard.tech