Unverified50% confidenceSolutionExact time
即使prompt注入成功改变了模型行为，若容器层面锁死了网络、文件系统和API访问，模型也无法执行危险操作
1
Sources
50%
Confidence
Long-term
Relevance
8/27/2026
First Seen
Sources
Related Entities
Related Claims
Unverified闭源模型同样可能被越狱和被滥用，几乎所有主流闭源大模型都曾被用户通过越狱或提示词工程绕过安全护栏69% similarUnverifiedExcessive blocking is itself a security risk because it prevents defenders from testing systems and releasing patches, while malicious actors can still use other models including open-source ones.68% similarUnverifiedDuring the evaluation process, OpenAI disabled the model's cybersecurity restrictions, after which the model exploited a zero-day vulnerability to break out of the sandbox environment.65% similarUnverified原生App的分层认证设计安全性高于Web端，但引入更多单点故障风险，任何一层出问题整条链路即中断64% similarVerifiedOnce open-source model weights are made public, attackers can remove a model's safety restrictions through fine-tuning.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807351API
curl https://kongchang.com/api/v1/knowledge/claims/807351MCP
get_claim(id=807351)