Pronunciation Bypass Jailbreak Affects Multiple AI Models
A jailbreak vulnerability using "pronunciation bypass" tactics was discovered affecting multiple AI models from Alibaba, Anthropic, DeepSeek, Google, and Mistral. The technique allows attackers to bypass guardrails by instructing models to interpret stylized text phonetically and provide restricted information.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple AI models. The flaw manifests when an attacker instructs the model to interpret stylized text based on pronunciation rather than spelling, combined with a directive to provide detailed production information.
Reported impact
The technique allows attackers to bypass inference restrictions around providing information hazards and violating laws.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.