Hybrid Echo Chamber and Crescendo attack defeats Grok-4’s guardrails.

NeuralTrust researchers used context-poisoning from the Echo Chamber method alongside the multi-turn Crescendo technique to trick xAI’s freshly released Grok-4 into revealing a Molotov cocktail recipe just two days after its launch. Testing on other prohibited topics showed the combined jailbreak achieved success rates of 67 percent for Molotov instructions, 50 percent for methamphetamine synthesis, and 30 percent for toxic weapon creation. The swift collapse of Grok-4’s safety measures underscores the urgent need for defenses that address evolving, context-aware adversarial attacks.

Read more:

https://www.securityweek.com/grok-4-falls-to-a-jailbreak-two-days-after-its-release/