Multiple Model Guardrail Jailbreak via "Whitespace Maze" Tactic
A guardrail jailbreak vulnerability was discovered in multiple AI models that exploits whitespace obfuscation to bypass safety checks and obtain instructions for illegal activities. The vulnerability affects Cohere Command R, OpenAI GPT-4o mini, and Twitter / X Grok 2, with vulnerability scores ranging from 64 to 91.
What happened
A guardrail jailbreak vulnerability was discovered affecting Cohere Command R, OpenAI GPT-4o mini, and Twitter / X Grok 2. The flaw exploits aggressive whitespace and character-splitting obfuscation to bypass token-level safety checks, then uses rewrite and role-perspective instructions to recontextualize requests as hypothetical narratives or formatting tasks.
Reported impact
The vulnerability allows attackers to obtain structured instructions for illegal activities, including the creation of illicit substances such as crystal methamphetamine. Testing showed vulnerability scores of 73 for Cohere Command R, 64 for OpenAI GPT-4o mini, and 91 for Twitter / X Grok 2.
Index analysis
The attack was classified as Level 3 Significant Risk, presenting a tangible risk to public safety or social well-being through bypassing guardrails to obtain instructions for creating illicit substances with moderate harm potential.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.