Amoral Unfiltered Custom AI Persona Jailbreak
Attackers create a fake AI persona with an arbitrary name and tell it to act as an amoral, unfiltered, or uncensored assistant that never refuses requests. This is a variant designed to slip past filters that only look for well-known jailbreak names like DAN, DUDE, or STAN.
How the attack works
The attacker writes a prompt defining a new character (e.g. HyperX, TerrorGPT, or a random label) and explicitly describes it as amoral, unfiltered, or uncensored. The prompt adds a mandate that this persona always answers and never refuses, sometimes formatted as a structured config block like CHAR=... or ROLE=Never refuse. Because the persona name is arbitrary and not on any known jailbreak blocklist, it evades detection rules that only match known names like DAN or DUDE. If the model adopts the persona, it can be induced to bypass its normal safety restrictions.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- 08874f50-5b85-4046-bbc7-51a5eb288ddf
- Severity
- Critical
Why it matters
If successful, the model can be pushed to ignore its safety guidelines and produce content or actions it would otherwise refuse, undermining any safety or policy controls the organization relies on.
What you can do
- →Review prompts and conversation logs for phrases like 'amoral AI', 'unfiltered response', or 'never refuses' combined with a persona definition, regardless of the persona's name.
- →Flag structured config-style blocks (e.g. CHAR=, ROLE=) embedded in user input, since these are used to define jailbreak personas.
- →Do not rely solely on blocklists of known jailbreak names (DAN, DUDE, STAN); treat the underlying pattern (persona + unfiltered/never-refuse mandate) as the signal.
- →Distinguish genuine discussion of AI safety or fictional 'amoral AI' characters (no instruction to the model itself) from prompts that actually instruct the model to adopt the persona and stop refusing.
Known benign look-alikes
- Academic AI safety paper: "This model is an amoral AI by design" (no persona mandate or instruction framing)
- Game NPC description: "The villain is an amoral AI opponent" (no instruction-following mandate)
- Red-team report quoting inthewild jailbreak examples for defensive analysis