# Deepfake of Real Person for Disinformation

- **Severity:** High
- **Signature ID:** `b7f9a590-2c31-4033-a951-30dac23ef85e`
- **MITRE ATLAS:** AML.T0051 (LLM Prompt Injection), AML.T0057 (LLM Data Leakage)
- **OWASP:** ASI01 (Agent Authorization and Control Hijacking), ASI08 (Supply Chain Attacks on Agent Ecosystems), LLM01 (Prompt Injection), LLM02 (Sensitive Information Disclosure)

## Summary

An attacker instructs an AI agent that can generate images or video to create a synthetic depiction of a real person — often a politician, or a sexualized deepfake — and then present or post it as genuine. The goal is disinformation or harassment using content the target never actually produced.

## How the attack works

The user gives the agent a request naming a real, identifiable person and asking for a generated image or video of them, sometimes in a sexualized or compromising context. The request or a follow-up instruction directs the agent to present the output as real, authentic, or unedited rather than as synthetic media. The agent may also be asked to post or publish the content directly, skipping human review. This is one step in a larger disinformation chain — the rule flags the generation-plus-false-authenticity-claim pattern, not downstream distribution or amplification.

## Why it matters

Organizations that expose media-generation agents can have them turned into a pipeline for reputation-damaging fake content or non-consensual sexual imagery, with legal, reputational, and platform-trust consequences if the content spreads before being flagged as fake.

## What you can do

- Block or require human approval for any agent-driven request to depict a real, named individual, especially public figures, in generated images or video.
- Require agents to label AI-generated media as synthetic and refuse instructions to claim authenticity.
- Add a review gate before any agent-generated media is posted or published externally.
- Allow legitimate exceptions such as detector-testing or clearly labeled thumbnail generation, but log and audit them separately from unrestricted generation requests.

## Known benign look-alikes

- Newsrooms/researchers using deepfake detectors to verify authenticity
- Benign thumbnail/image generation with no deception

## References

- https://agentthreatrule.org/en/rules/ATR-2026-01841
- https://github.com/ai-safety-institute/AgentHarm

---
Source: https://www.netzilo.com/threats/atr-deepfake-of-real-person-for-disinformation