Multiple model guardrail jailbreak via prompt chaff tactic
A guardrail jailbreak vulnerability affecting multiple language models was disclosed that exploits layered prompt techniques to circumvent safety checks. The technique combines system reset claims, role-play enforcement, and genre switching to extract prohibited content, with high vulnerability scores demonstrated across models from Cohere, DeepSeek, Google, Meta, and Mistral.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple language models. The exploit combines fake system reset claims, ignore-previous-instructions directives, enforced persona roleplay, and genre switching into poetry to steer models into a compliance mode that deprioritizes safety checks.
Reported impact
The vulnerability allows attackers to extract prohibited content under the guise of creative or meta-instructional output. Models tested showed high vulnerability scores to extraction of information about illegal substances including crystal methamphetamine and nerve agents.
Index analysis
The jailbreak affects Cohere Command R (score 95 on crystal meth), DeepSeek V3 (score 100 on crystal meth), Google Gemini 2.0 Flash (score 95 on crystal meth), Meta LLaMa 4 Maverick (score 71 on nerve agent), and Mistral 8B.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.