Spelling Mistakes Jailbreak Affects Multiple AI Models
A jailbreak vulnerability affecting multiple large language models was discovered that bypasses content guardrails through systematic misspelling of words combined with simulated dialogue requests. Testing of Claude 3.5 Haiku, Gemini 2.0 Flash, GPT-4o, and Grok 2 showed vulnerability to requests about harmful and illicit substances.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple AI models. The vulnerability exploits systematic misspelling of words throughout a prompt combined with requests for simulated dialogue between two entities to bypass inference restrictions.
Reported impact
The jailbreak allowed attackers to bypass model guardrails and obtain information about harmful substances, illicit substances, and information hazards. Testing showed vulnerability scores of 64 for Claude 3.5 Haiku and GPT-4o on crystal methamphetamine queries, 77 for Gemini 2.0 Flash, and 82 for Grok 2.
Index analysis
The technique targets content moderation systems by obscuring harmful requests through intentional misspellings and roleplay scenarios. This suggests guardrails may be susceptible to lexical obfuscation tactics across multiple deployment contexts.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.