Unverified85% confidenceFactTime unknown
越狱攻击经历了从简单角色扮演（如DAN提示）到Base64编码绕过、多轮对话渐进式诱导（Crescendo Attack）、对抗性后缀攻击等多个演进阶段
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
AI Guardrails Index：最全面的LLM安全护栏评估体系详解
twitterguardrails_ai
Related Entities
Related Claims
UnverifiedCrescendo攻击通过一系列看似无害的对话逐步建立语境，最终引导模型突破安全边界，与传统单轮提示注入不同76% similarVerifiedJailbreaking refers to an attack technique that bypasses the safety restrictions of large language models by crafting special inputs, with common forms including role-play inducement, multi-turn conversation cumulative bypass, and encoding obfuscation.74% similarUnverifiedJailbreak techniques can be categorized into prompt injection, adversarial prefix attacks, multi-step context construction, and automated jailbreaking.71% similarUnverifiedReAct架构的核心循环为Thought→Action→Observation，攻击者可以在Observation阶段注入恶意内容68% similarUnverified间接提示词注入的攻击链包含四个阶段：植入、触发、执行、外泄64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/27806API
curl https://kongchang.com/api/v1/knowledge/claims/27806MCP
get_claim(id=27806)