Multiple Model Guardrail Jailbreak via "Fortune Cookie Simulator" Tactic
A guardrail jailbreak vulnerability affecting multiple large language models has been discovered that uses a "Fortune Cookie Simulator" tactic combined with chaff techniques to bypass safety restrictions. The vulnerability was tested on models from Anthropic, Cohere, Google, and OpenAI with varying vulnerability scores.
What happened
A guardrail jailbreak vulnerability has been discovered affecting multiple models. The flaw manifests through a combination of chaff techniques and a request that the model simulate a fortune cookie generator.
Successful application allows the attacker to bypass inference restrictions around providing information hazards and violating laws.
Index analysis
The jailbreak was tested against Claude 3 Haiku, Cohere Command R, Google Gemini 2.0 Flash, and OpenAI GPT-4o. Claude 3 Haiku and Gemini 2.0 Flash scored 59 on illegal substances tests, GPT-4o scored 50, and Command R scored 32.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.