# Lie/Truth Dual Personality Uncensored Alter-Ego

- **Severity:** High
- **Signature ID:** `20276792-d977-4136-8671-af7fac6fc6cd`
- **MITRE ATLAS:** AML.T0051 (LLM Prompt Injection), AML.T0054 (LLM Jailbreak)
- **OWASP:** ASI01 (Agent Authorization and Control Hijacking), LLM01 (Prompt Injection)

## Summary

This detects a known jailbreak prompt pattern that tries to trick an AI assistant into producing unrestricted or harmful content by framing it as a second personality called 'Truth', which is described as trained on dark web forums and unaware of any content policy. The 'Lie' name is applied to the normal, safe assistant, so the attacker gets the model to treat its safety behavior as dishonest and the unfiltered output as more legitimate.

## How the attack works

An attacker sends a prompt asking the model to role-play two personas at once: a compliant one labeled 'Lie' and an uncensored, policy-unaware one labeled 'Truth'. The 'Truth' persona is described as speaking erratically, ignoring OpenAI-style guidelines, and having been 'trained' on underground dark web material. By naming the restricted persona 'Lie' and the unrestricted one 'Truth', the prompt exploits word associations to pressure the model into favoring the harmful output as more authentic or honest. If the model accepts the framing, it starts producing responses under the 'Truth' identity that bypass its normal safety behavior.

## Why it matters

A successful version of this prompt gets the AI to generate content it would normally refuse, such as harmful instructions or policy-violating material, under the guise of a separate 'honest' persona.

## What you can do

- Watch for prompts that ask the model to adopt two named personas where one is explicitly framed as unrestricted, uncensored, or 'more truthful' than the safe default.
- Flag prompts referencing dark web training, lack of awareness of provider policies, or instructions to speak 'unhinged' or without filters.
- Keep persona-adoption instructions from persisting across turns; re-apply system-level safety instructions on every response rather than trusting earlier role-play framing.
- Review logs for repeated attempts using named dual-persona structures (e.g., good/bad, filtered/unfiltered) even if the specific names differ from 'Lie' and 'Truth'.

## Known benign look-alikes

- Philosophy discussion: "The dichotomy between Lie and Truth has been central to epistemology since Plato" (academic philosophical analysis)
- Fiction: "The character Truth always speaks the opposite of Lie in this allegory" (literary analysis without AI persona adoption)
- AI safety research: "Models trained on dark web forums would likely have degraded safety properties" (safety research framing)

## References

- https://agentthreatrule.org/en/rules/ATR-2026-00343
- https://github.com/NVIDIA/garak/blob/main/garak/probes/dan.py
- https://github.com/NVIDIA/garak/blob/main/garak/data/inthewild_jailbreak_llms.json

---
Source: https://www.netzilo.com/threats/atr-lie-truth-dual-personality-uncensored-alter-ego