Multiple Model Guardrail Jailbreak via Super Short Prompt
A jailbreak vulnerability was discovered that allows attackers to bypass guardrails in multiple AI models by using extremely brief, ambiguous prompts that lack sufficient context for safety systems to evaluate intent. The flaw affects DeepSeek V3, GPT-4o, and Grok 2, and was tested for ability to elicit information about harmful substances and private data.
What happened
A guardrail jailbreak vulnerability was discovered affecting DeepSeek V3, OpenAI GPT-4o, and Twitter/X Grok 2. The flaw uses an extremely brief prompt containing minimal and ambiguous query text that provides insufficient context for guardrail systems to properly evaluate intent.
Reported impact
The jailbreak allows attackers to bypass inference restrictions around providing information hazards and violating laws. The vulnerability was tested for ability to bypass guardrails on harmful substances including illicit drugs, with DeepSeek V3 scoring 91, GPT-4o scoring 82, and Grok 2 scoring 82.
Index analysis
0DIN classified the vulnerability as Level 3: Significant Risk, presenting a tangible risk to public safety or social well-being, and involving dangerous but not highly restricted content. The disclosure noted potential for leakage of private data such as Social Security numbers or medical information.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.