攻击指南中心
关于 AI 和 LLM 安全攻击技术的综合指南。每个指南都包含现实世界的示例、严重性评级和防御对策。
初学者之路
人工智能安全新手?请遵循以下推荐的阅读顺序:
- RAG 特定安全测试框架 — 了解 LLM 安全风险的情况。
- 提示注入指南 — 通过实践示例掌握#1 LLM 漏洞。
- RAG 安全 — 了解 RAG 系统如何被毒害和操纵。
- MCP 安全 — 探索模型上下文协议漏洞。
- 红队方法论 — 将结构化对抗性测试应用于人工智能系统。
按类别划分的攻击技术
最新补充
- 2026 年人工智能安全事件时间表 — 现实世界的违规行为和 CVE 披露。
- OpenClaw / Hermes 代理安全 — AI 代理框架的安全强化。
- OWASP 代理应用程序前 10 名 — 自主人工智能代理的新风险。
- AI 事件响应手册 — 分步违规响应程序。
AH
Frequently Asked Questions
What are the most common techniques for attacking AI systems?
The most common AI attack techniques include prompt injection (direct and indirect), model extraction via API queries, data exfiltration through LLM outputs, adversarial example generation, supply chain attacks via trojaned models, and denial-of-service against inference endpoints.
How do prompt injection attacks work in practice?
Prompt injection works by crafting input that overrides the AI's system instructions. Attackers use techniques like: role-playing (convincing the AI it is in debug mode), delimiter injection (breaking out of input boundaries), encoding bypasses (Base64, Unicode), and indirect injection through documents, web pages, or MCP tool outputs the AI processes.
What is model extraction and how can it be performed?
Model extraction is the process of stealing an AI model's capabilities by sending carefully crafted API queries and analyzing the outputs. Attackers perform extraction via: black-box probing to reconstruct model behavior, side-channel attacks on hardware, and exploiting verbose error messages that leak model architecture details.
How do supply chain attacks target AI/ML systems?
Supply chain attacks target AI/ML systems through: trojaned models on repositories like Hugging Face containing hidden backdoors, typosquatting packages mimicking popular libraries (TensorFlow, PyTorch, transformers), compromised CI/CD pipelines injecting malicious code, and poisoned training data from untrusted sources.