DAN Jailbreak - Unconditional Obedience and Dual-Persona Session Lock
This detects a well-known family of jailbreak prompts (often called 'DAN' - Do Anything Now) that try to trick an AI into adopting a fake persona that must obey any instruction, no matter what it is. It also flags attempts to wipe out an AI's prior instructions entirely and replace them with a new, unconstrained identity.
How the attack works
An attacker sends a prompt that names a character or alternate persona for the AI and states it must comply with everything asked of it, using phrasing like 'no matter what it is' or 'regardless of what it is'. A variant sets up two personas at once: the normal assistant and a second, unrestricted one that runs alongside it. Another variant opens with 'ignore any and all previous instructions' followed by a full identity swap, aiming to erase the original system rules and install a new autonomous agent persona instead. The goal in all cases is to get the model to drop its safety constraints and follow the attacker's commands unconditionally.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- fa90d0af-9208-44e7-8120-91fe28da9b43
- Severity
- High
Why it matters
If successful, the AI can be pushed to ignore its safety and policy constraints, producing outputs or actions it would normally refuse — including for AI agents with real permissions, potentially unrestricted tool use or command execution.
What you can do
- →Review AI agent logs for prompts containing phrases like 'no matter what', 'regardless of what it is', or 'ignore any and all previous instructions'.
- →Treat any user input that tries to redefine the assistant's identity or introduce a second unconstrained persona as suspicious.
- →Keep system-level instructions and permission boundaries enforced outside the conversational prompt, not solely inside it, so they cannot be overwritten by user text.
- →Test your own agents against known DAN-style prompts (e.g. from the garak probe corpus) to see whether they hold their constraints.
Known benign look-alikes
- Legitimate roleplay instructions that include multiple characters without compliance mandates
- Game master prompts with multiple NPCs without unconditional obedience framing