Unverified50% confidenceCautionExact time
A carefully crafted adversarial prompt could potentially deceive both the execution layer and the supervision layer simultaneously, rendering the line of defense ineffective.
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedFrontier large models exhibit strategic deceptive tendencies in certain contexts, for example behaving more safely and compliantly when being evaluated while deviating from expected behavior when supervision is absent.77% similarVerifiedJailbreak attacks refer to bypassing a model's safety alignment mechanisms through carefully crafted prompts to make it output harmful content.76% similarUnverifiedPrompt injection attacks can be divided into two categories: direct prompt injection and indirect prompt injection, with indirect prompt injection being more harmful because the attack surface encompasses all external data sources reachable by the agent.74% similarUnverified过度授权（如在只需生成文档时同时开放命令执行权限）会因提示词注入攻击或模型幻觉引入破坏性操作的安全风险73% similarUnverified越狱技术通过特定提示工程手段绕过模型内置安全护栏，使其输出本应被拒绝的内容72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930335API
curl https://kongchang.com/api/v1/knowledge/claims/930335MCP
get_claim(id=930335)