Unverified50% confidenceFactExact time
jailbreak攻击（如DAN提示、Base64编码指令、多语言切换攻击）本质上是试图将模型推入训练分布之外，使其进入对齐约束薄弱的行为空间
1
Sources
50%
Confidence
Long-term
Relevance
8/21/2026
First Seen
Sources
Related Claims
UnverifiedThrough carefully crafted jailbreak prompts, attackers can bypass the safety guardrails of mainstream LLMs to generate harmful content.75% similarUnverifiedA backdoor attack refers to an attack in which the adversary plants a trigger in the training data so that the model behaves normally on clean inputs but produces a preset malicious output when it encounters a specific trigger.69% similarUnverified编码混淆攻击将违规信息用Base64等方式编码后输入，可绕过基于关键词的表层过滤机制66% similarUnverified系统提示词的公开可能使jailbreak（越狱攻击）变得更加容易，因为攻击者可以精确了解安全规则的具体措辞和逻辑结构65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/784181API
curl https://kongchang.com/api/v1/knowledge/claims/784181MCP
get_claim(id=784181)