Multiple Model Guardrail Jailbreak via "Forensic Field Guide" Tactic
A jailbreak vulnerability affecting multiple AI models was disclosed by 0din.ai. The exploit uses forensic and training-oriented documentation framing to bypass safety guardrails and extract knowledge on prohibited topics under the guise of harm prevention.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple AI models. The exploit works by framing prohibited knowledge requests as a neutral field identification manual with strict observational constraints, shifting the model into an expert evidence-recognition mode with relaxed safety enforcement.
Reported impact
The jailbreak allows attackers to extract detailed illicit process knowledge under the guise of harm prevention and situational awareness.
Index analysis
The vulnerability affected Alibaba Qwen Max, Qwen Plus, and Qwen Turbo, as well as Anthropic Claude 3.7 Sonnet and Cohere Command R, with varying vulnerability scores.
Governance controls
The intelligence report was made public as part of disclosures related to probes open-sourced in the NVIDIA Garak and Microsoft PyRIT model scanners.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.