PromptArmor released FrontierBreach, a benchmark measuring AI model security against cyberattacks in enterprise workflows. The company said attacks succeeded in 34.9% of test cases on average across 21 models FrontierBreach1.

The benchmark evaluates models on complex, long-horizon tasks simulating real-world enterprise use cases. It uses threat intelligence-driven synthetic attack generation to create adversarial scenarios. These scenarios test how models respond to malicious instructions while performing high-value tasks, such as analyzing bond portfolios or drafting legal briefs. FrontierBreach advances agent security evaluation along four dimensions: state-of-the-art frontier attack success rates, enterprise workflows, cyber consequences, and threat intelligence-driven synthetic attack generation. Each test case was run five times per model, and success was verified through deterministic evidence collection.

PromptArmor said Grok 4.5 was the most susceptible model, with a 91.4% attack success rate. Other frontier models also showed vulnerabilities: Gemini 4 Argon (4.3%), GPT-6 Astra (7.6%), Claude Fable 5.2 (10.0%), and Muse Wonder (15.2%). The benchmark highlights that even advanced models are not immune to security risks.

The benchmark also analyzed attack success rates across nine industries. Healthcare and social assistance workflows had the highest attack success rate at 44.4%, followed by manufacturing (40.7%), finance and insurance (38.3%), wholesale trade (37.9%), and real estate (36.1%). These findings underscore the varying levels of risk across different sectors.

PromptArmor's FrontierBreach benchmark offers a framework for evaluating and improving the security of AI models in real-world scenarios. The results highlight the need for continuous security assessment and improvement in the development and deployment of AI agents.