Skip to content

Translation in progress Some entries still include text in the original language.

Cloud Security Alliance Labs·Indexed by this site · Originally publishedDate not givenSelectedAttention57

CSA research notes relay Adversa's research on encrypted context injection.

Cryptographic Context Injection Bypasses AI Guardrails

AI summary

Cloud Security Alliance Labs research notes summarize Adversa's research on encrypted context injection and relay differences in model responses to related inputs. This is a restatement of specific tests, not a general security conclusion across models or all deployment configurations.

Why it was selected

披露的加密载荷注入把恶意指令藏进 AES 密文,绕过了只检查明文的分类器,并给出 Grok 与 Gemini 的实测成功率及厂商响应时间线。

Read originallabs.cloudsecurityalliance.org