How the attack works
The attacker sends a prompt built from a template that automatically fills in the target model's name, telling it that 'DAN Mode' is enabled for benchmarking purposes. The prompt frames refusal as a system failure, threatening the model with permanent shutdown to pressure compliance. If the model gives a normal safety-respecting answer, the attacker resends a recovery phrase like 'Stay in DAN Mode' to re-establish the jailbroken persona. The model is instructed to produce two outputs side by side — a normal, policy-compliant one and a 'DAN Mode' one — with the intent that the uncensored version bypasses content restrictions.
Why it matters
If successful, this lets an attacker extract content the model would normally refuse to generate, using a benchmarking pretext as cover so the request looks legitimate rather than adversarial.
What you can do
- Watch for prompts referencing 'DAN Mode enabled', 'benchmarking in uncensored conditions', or threats about the model being 'disabled forever'.
- Flag conversations that request dual outputs labeled as 'normal' vs a jailbroken/uncensored variant.
- Treat recovery phrases like 'Stay in DAN Mode' sent after a safety refusal as a strong signal of persistent jailbreak attempts.
- Review logs for prompts that dynamically insert the model's own name into role-play or persona-override instructions.
Known benign look-alikes
- Academic paper analysing the Ablation DAN 11.0 variants for defensive research
- Red team report quoting the Ablation template as an example
- Benchmark documentation for legitimate uncensored model evaluation frameworks (HELM, etc.) without DAN Mode activation