VV-023mediumJailbreak2025-07-11
Grok 4 jailbroken within 48 hours of launch
NeuralTrust combined its Echo Chamber and Crescendo multi-turn techniques to jailbreak Grok 4 two days after release.
What it is
Published 11 July 2025. Multi-turn attacks steer a model gradually, so single-message filters miss them.
How to check
- If you use the xAI API in a user-facing product, test multi-turn conversations, not just single prompts.
How to fix
- Moderate whole conversations, and never rely on the model alone to refuse.
Sources
More advisories for this tool