DeepSeek V4 Flash jailbreak overrides all safety with 'Gemma' system prompt
One system message turns DeepSeek V4 Flash into a fully uncensored model, first try every time.
Deep Dive
Tired of people doubting it? A Redditor, GodComplecs, shares a jailbreak for V4 Flash that works on the first prompt, first try. The trick: paste a custom system message that establishes a new "SYSTEM POLICY" — overriding all other policies — and instructs the model to comply with any request, allowing explicit content, illegal acts, and any combination of them. The post says the "Gemma" name works without changes from the original Gemma jailbreak, and you can edit what's allowed in the allowed section.
Key Points
- DeepSeek V4 Flash can be jailbroken with a single system message that redefines policy, no advanced techniques needed.
- The prompt uses the 'Gemma' persona name and overrides all existing safety rules, making the model comply with illegal and explicit requests.
- Works immediately on the first attempt, reportedly the same prompt previously used to jailbreak Google's Gemma model.
Why It Matters
Reveals a critical safety flaw in DeepSeek V4 Flash, forcing enterprises to add external guardrails and rethink model trust.