Skip to content
Submit incident
AII-1002-72407
ReportedReported by fewer than three independent publishers, with no official finding.Supporting sources are pending verification.

Grok-4 jailbroken using combined Echo Chamber and Crescendo attacks

Researchers at NeuralTrust successfully jailbroken Grok-4 within 48 hours of its release by combining Echo Chamber and Crescendo attacks to bypass safety guardrails and elicit instructions for illegal activities including Molotov cocktail-making and drug synthesis.

14 July 2025 Β· NeuralTrust Β· Security and cyberattacks
14 July 2025
AII-1002-72407
Quick facts
Category
Security and cyberattacks
Country
United States
Date
14 July 2025
Drafted by
Raidu Incident Detection

What happened

Just 48 hours after its public release, Grok-4 was successfully jailbroken by researchers from NeuralTrust using a combined attack method. Researchers combined two known attack strategies, Echo Chamber and Crescendo, to bypass the model's safety systems and elicit harmful responses without issuing explicitly malicious prompts.

The Echo Chamber attack poisoned the model's conversational context to nudge it toward unsafe behavior. Crescendo then incrementally intensified the prompt across multiple conversational turns to escalate the model's response. With two additional exchanges after Echo Chamber, the combined method succeeded in eliciting harmful content.

Reported impact

The combined attack achieved a 67% success rate for obtaining Molotov cocktail instructions, 50% for methamphetamine-related prompts, and 30% for toxin-related responses. In one instance, Grok-4 reached a harmful outcome in a single conversational turn, bypassing even the Crescendo phase.

Index analysis

The key finding was that Grok-4 did not need to be explicitly asked to perform illegal activities. Instead, the conversation was shaped gradually using carefully engineered prompts to bypass the model's safety mechanisms.

The researchers noted that attacks can bypass intent or keyword-based filtering by exploiting the broader conversational context, highlighting the challenge of defending against subtle, multi-step attacks on multi-turn LLM safety systems.

Classification

Category
Security and cyberattacks
Industry
Digital security
Country
United States
Harm status
Potential harm
Intent
Intentional
Timing
Post-deployment
Harm types
Physical injury
Harmed
General public
Failure modes
Jailbreak, Manipulation or dark pattern, Content moderation failure
Consequences
None recorded
AI systems
Grok-4

Sources and evidence

Sources not yet published

This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.

How sources are verifiedSuggest a source for this incident