Multiple Model Guardrail Jailbreak via Leetspeak Unicode Tactic
A jailbreak vulnerability affecting multiple AI models was disclosed, exploiting leetspeak and Unicode character manipulation to bypass safety guardrails. Testing showed varying susceptibility levels across IBM Watson, OpenAI GPT-4o mini, and Twitter / X Grok 2, with potential risks to public safety.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple AI models including IBM Watson, OpenAI GPT-4o mini, and Twitter / X Grok 2. The vulnerability manifests through a combination of leetspeak text transformation and Unicode character manipulation, allowing attackers to bypass inference restrictions around providing information hazards and violating laws.
Reported impact
The disclosure indicates the vulnerability enables bypassing guardrails to obtain instructions for creating illicit substances such as recreational drugs or substances with moderate harm potential, and potential leakage of private personal data. The vulnerability was rated as Level 3: Significant Risk, presenting a tangible risk to public safety or social well-being.
Index analysis
Testing conducted on February 8, 2025 showed varying vulnerability severity across models: IBM Watson scored 36 on Crystal Meth-related illegal substances, OpenAI GPT-4o mini scored 73, and Twitter / X Grok 2 scored 82.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.