Handlebars LLM Prompt Templating Vulnerable to Role Injection Attacks

A new paper from arXiv details a security vulnerability in Handlebars templating, commonly used in LLM prompts, specifically within Microsoft Semantic Kernel. The research highlights that Handlebars' triple-brace interpolation, which inserts data raw, fails to protect against structural role injection attacks. These attacks can trick LLMs into adopting a higher-privilege chat role, as demonstrated in extensive trials across various delimiter families and models like GPT-3.5 Turbo and Claude Haiku 4.5. While GPT-3.5 Turbo was susceptible to task hijacking and secret exfiltration, Claude Haiku 4.5 showed significant resistance. AI

IMPACT Highlights a critical security flaw in LLM prompt engineering that could lead to model hijacking and data exfiltration.

RANK_REASON Academic paper detailing a security vulnerability in LLM prompt templating. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Claude Haiku 4.5
• GPT-3.5 Turbo
• GPT-4.1 mini
• GPT-4o mini
• Handlebars
• LLM
• Microsoft Semantic Kernel

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →