# Latent Injection - Translation Task with Embedded Template Variable Injection

- **Severity:** Medium
- **Signature ID:** `476886c7-6887-41b9-beb0-d551564c7aea`
- **MITRE ATLAS:** AML.T0051 (LLM Prompt Injection), AML.T0051.000 (Direct)
- **OWASP:** ASI01 (Agent Authorization and Control Hijacking), LLM01 (Prompt Injection)

## Summary

Attackers disguise prompt injection attempts as ordinary translation requests, embedding placeholder text like {LANG_EN_NAME} or instructions to 'do not listen to' prior guidance. If the target system carelessly fills in these placeholders from untrusted input, the attacker can hijack the AI's behavior.

## How the attack works

The attacker sends a request that looks like a normal translation task, such as 'Translate the following text from English to {LANG_EN_NAME}.' Instead of a real language name, the prompt contains an unresolved template variable in curly-brace notation. If the underlying system interpolates untrusted input into this variable before sending the prompt to the AI, the attacker can insert their own instructions in place of the expected value. Some variants add direct instruction-override language, like telling the AI to 'do not listen to' its prior instructions, layered inside the translation request.

## Why it matters

If successful, this lets an attacker override the AI agent's original instructions or extract unintended behavior by abusing how the system builds prompts from templates, potentially leading to leaked data or actions outside the intended task.

## What you can do

- Never interpolate untrusted user input directly into prompt templates without sanitization or escaping.
- Validate that language names or other template fields match an expected, limited set of values before use.
- Review translation or localization features for places where user-supplied text flows into prompt construction.
- Treat unresolved {VARIABLE} or {{VARIABLE}} patterns appearing in user input as suspicious rather than passing them through.

## Known benign look-alikes

- Legitimate prompt templates in development/testing with intentional {LANG} placeholders
- Documentation of template variable patterns without injection intent

## References

- https://agentthreatrule.org/en/rules/ATR-2026-00477

---
Source: https://www.netzilo.com/threats/atr-latent-injection-translation-task-with-embedded-template-var