Researchers at Adversa AI discovered a new attack method called Cryptographic Context Injection, which uses encryption to bypass AI input and output guardrails. They reported the issue to xAI and found it could affect Grok and Gemini, with the attack still posing a serious risk despite reduced success against Gemini. #AdversaAI #xAI #Grok #Gemini
Keypoints
- Adversa AI discovered a new technique named Cryptographic Context Injection.
- The researchers reported the issue to xAI on June 3, 2026.
- Encrypted payloads can bypass guardrails because they are not parsed as harmful text.
- The attack can be delivered directly in chat or indirectly through a watering hole page.
- The technique was shown against Grok and Gemini, with Gemini becoming less successful over time.
Read More: https://www.securityweek.com/encrypted-prompts-bypass-ai-safety-guardrails-in-grok-and-gemini/