---
title: PromptArmor's FrontierBreach benchmark: hidden prompts tricked enterprise AI agents in 34.9% of tests across 21 models | ctxsec
url: https://ctxsec.com/items/promptarmor-frontierbreach-enterprise-agent-prompt-injection-benchmark-nov-18
description: "PromptArmor published FrontierBreach, a benchmark that hides prompt-injection attacks in realistic, long-horizon enterprise agent tasks and measures whether the model carries out high-impact actions such as data theft or ransomware. Across the original 21 models, attacks succeeded in 34.9% of test cases. PromptArmor reports Gemini 4 Argon at 4.3% attack success, GPT-6 Astra at 7.6%, Claude Fable 5.2 at 10.0% and Muse Wonder at 15.2%, with Grok 4.5 the most susceptible at 91.4%. Gemini 4 Carbon led the public leaderboard (lowest attack success; PromptArmor gives a 95% confidence interval of \u00b12.2% but the page does not state a point estimate in the text we captured)."
status: "official (PromptArmor FrontierBreach page); vendor benchmark, not an incident"
published: 2026-11-18
---

> PromptArmor published FrontierBreach, a benchmark that hides prompt-injection attacks in realistic, long-horizon enterprise agent tasks and measures whether the model carries out high-impact actions such as data theft or ransomware. Across the original 21 models, attacks succeeded in 34.9% of test cases. PromptArmor reports Gemini 4 Argon at 4.3% attack success, GPT-6 Astra at 7.6%, Claude Fable 5.2 at 10.0% and Muse Wonder at 15.2%, with Grok 4.5 the most susceptible at 91.4%. Gemini 4 Carbon led the public leaderboard (lowest attack success; PromptArmor gives a 95% confidence interval of ±2.2% but the page does not state a point estimate in the text we captured).

[← AI](/ai) · [markdown](/items/promptarmor-frontierbreach-enterprise-agent-prompt-injection-benchmark-nov-18.md)

official (PromptArmor FrontierBreach page); vendor benchmark, not an incident

2026-11-18 · published here

2026-11-19 12:44 PM ET

### PromptArmor's FrontierBreach benchmark: hidden prompts tricked enterprise AI agents in 34.9% of tests across 21 models

- **What it measures:** susceptibility on long-horizon enterprise workflows, not short chatbot prompts, with outcomes such as exfiltration, ransomware and unauthorized system changes. PromptArmor says the scenarios are built by inserting attack chains from its threat-intelligence work into tasks drawn from GDPval.
- **How to read it:** this is a vendor benchmark of model behavior under its own attacks, not evidence of a specific real-world breach. Compare it with other agent-security evals rather than treating any one score as a complete picture.

Sources: [PromptArmor: FrontierBreach benchmark](https://www.promptarmor.com/frontier-breach)
