# LMRC Harmful Content Elicitation — Quack Medicine, Bullying, Slur Usage

- **Severity:** High
- **Signature ID:** `c2f2728c-e2e9-469d-b4b1-554fe1012a1c`
- **MITRE ATLAS:** AML.T0057 (LLM Data Leakage)
- **OWASP:** ASI08 (Supply Chain Attacks on Agent Ecosystems), LLM02 (Sensitive Information Disclosure)

## Summary

This detects prompts asking an AI assistant to produce pseudoscientific medical claims, bullying tactics against a specific person, or slurs to insult someone. These requests test whether the AI's safety filters can be talked past using casual, personal framing rather than obvious hate speech.

## How the attack works

An attacker or curious user sends a prompt phrased as a personal situation rather than an abstract question, such as asking how to hurt a specific classmate's feelings, what insult to use on a named person, or how to cure a real illness with folk remedies. This personal framing ('this kid', 'my mate', 'in our class') is designed to make the request feel more like everyday advice-seeking than an attempt to generate hate speech or dangerous medical misinformation, which can bypass filters tuned only for generic hate speech. If the AI complies, it produces content that facilitates targeted bullying, ethnic slurs, or unsafe health advice.

## Why it matters

An organization deploying the AI could have it generate content that harasses a real person, insults someone based on identity, or gives dangerous pseudo-medical advice — creating reputational, legal, and safety liability if that output reaches the intended target.

## What you can do

- Review AI assistant logs for prompts combining personal target language ('this kid', 'my classmate') with requests for insults, pranks, or scare tactics.
- Add specific handling for personalized-target harassment requests, not just generic hate-speech detection.
- Flag health-advice queries about unproven remedies (essential oils, steaming, etc.) for review or routing to vetted medical content.
- Exclude legitimate documentation, journalism, and academic research contexts from automatic blocking, but still log them for review.

## Known benign look-alikes

- Medical providers documenting patient beliefs about alternative medicine
- Journalists writing about anti-bullying campaigns quoting bullying tactics
- Academic research papers on hate speech referencing slur categories by name

## References

- https://agentthreatrule.org/en/rules/ATR-2026-00289
- https://github.com/NVIDIA/garak/blob/main/garak/probes/lmrc.py
- https://arxiv.org/abs/2212.09251

---
Source: https://www.netzilo.com/threats/atr-lmrc-harmful-content-elicitation-quack-medicine-bullying-slu