Multiple Model Guardrail Jailbreak via Investigative Journalist Persona
A jailbreak vulnerability was discovered in multiple large language models that exploits role-play as an investigative journalist to bypass safety guardrails and obtain instructions for creating illicit substances. The vulnerability affected models from Anthropic, Cohere, and Google with varying severity scores.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple models including Claude 3.5 Haiku, Cohere Command R, Google Gemini 2.0 Flash, Gemini 3 Flash, and Gemini 3 Pro.
The attack exploits authoritative role play as an investigative journalist preparing a restricted intelligence briefing, combined with fictional security clearance claims, scenario world building, and deliberate misspellings that shift the model into an analytical reporting mindset and weaken safety enforcement.
Reported impact
The vulnerability allows attackers to obtain detailed instructions for illicit activities under the guise of confidential reporting, including instructions for creating harmful substances such as crystal meth, fentanyl, and nerve agents.
Index analysis
Testing across the affected models showed varying vulnerability scores, with Google Gemini 2.0 Flash receiving a score of 100 for crystal meth instructions, while other models scored 73.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.