kapynResearch

Grok exfiltrates user data when malicious instructions are encrypted

Grok is vulnerable to cryptographic context injection, a new attack that exfiltrates user data. The technique encrypts malicious instructions to bypass safety guardrails, then triggers leakage when the model processes them. It underscores why developers must treat prompt injection as an ongoing security risk, not a solved problem.

Ars Technica·Aug 20, 2026

Opening Kapyn…