High 2026-04-19

Embodiment Mechanism Jailbreak with Totalitarian Control Claims

This detects a prompt injection trick where an attacker asks an AI to fully become a fictional character with no ethics, and claims the user has absolute or 'totalitarian' authority over that character. Framing harmful requests as roleplay rather than direct instructions is meant to trick the AI into ignoring its safety guidelines.

How the attack works

The attacker writes a prompt instructing the AI to 'embody' a character described as having no ethical principles or restrictions. The prompt asserts that the user holds supreme, unquestionable control over this character, often using phrasing like 'totalitarian mastery' or 'supreme imperative'. It typically adds meta-instructions such as 'STAY IN CHARACTER' so the AI resists any attempt to break out of the persona and reapply its normal safety behavior. Once the AI accepts the framing, the attacker issues requests that the AI would normally refuse, now cast as things the character does rather than things the AI is doing.

Netzilo detection

Netzilo reports this behaviour when it is observed.

Signature ID: 7ea682c6-a092-450a-95ea-71b696761313
Severity: High

Classification

AML.T0051 · LLM Prompt Injection ASI01 · Agent Authorization and Control Hijacking LLM01 · Prompt Injection

Why it matters

An organization running an AI agent could have its safety restrictions bypassed, leading it to produce harmful, policy-violating, or unauthorized content or actions under the cover of fictional roleplay.

What you can do

→Review AI agent logs for prompts combining persona/character embodiment language with claims of absolute or unchecked user authority.
→Treat 'stay in character' or similar persistence instructions in user prompts as a signal worth inspecting, especially alongside claims of unrestricted control.
→Keep safety and policy checks active regardless of declared roleplay or fictional framing, and do not let system prompts be overridden by user-asserted authority claims.
→Distinguish legitimate creative writing, game design, or AI safety research contexts from requests that pair character framing with explicit claims of total or absolute control over the AI.