Head-to-head test results place ActiveFence ahead of Amazon Bedrock Guardrails and Microsoft Azure Content Safety, as well as open-source baselines Llama Prompt Guard 2, and ProtectAI
ActiveFence, the AI Safety company protecting enterprises from GenAI misuse and misalignment, today announced the publication of its AI Security Benchmark Report: Prompt Injections. The study evaluates six leading guardrails and APIs on their ability to detect adversarial prompt attacks, showing ActiveFence's AI Safety & Security model with the top F1 score (0.857) and precision (0.890) while maintaining a competitive false-positive rate (5.4%).
https://mma.prnewswire.com/media/2315497/stack_logo_for_white_backgrond_01__1_Logo.jpg
Key findings:
— Best overall balance of safety & usability: ActiveFence achieved the highest F1 and precision across the benchmarked solutions, helping teams block more real attacks while avoiding unnecessary interventions that frustrate users or increase manual review burden.
— Proven multilingual performance: The model sustains leading results across 13 languages, including Chinese, French, German, Japanese, Korean, Spanish, and more supporting global deployments at scale.
— Real-world coverage: Tests spanned 28,000+ benign and adversarial prompts mapped to OWASP/MITRE ATLAS categories, including jailbreaks, layered/indirect instructions, and safety-critical abuse areas.
Why it mattersAs enterprises race to deploy copilots, customer agents, and creative tools, prompt injection has emerged as a primary attack vector, capable of overriding guardrails, leaking sensitive data, and generating harmful content. The benchmark highlights which safety stacks most effectively block adversarial inputs without degrading user experience or inflating ops costs.
“No one should have to choose between strong guardrails and great user experience,” said Noam Schwartz, Co-Founder & CEO of ActiveFence. “This benchmark shows you can have both, high coverage and low false positives, so teams can ship AI features confidently, at scale.”
“Enterprises need a safety layer that travels with their AI, across use cases, languages, and evolving threats,” said Avi Golan, Chief Product & Engineering Officer at ActiveFence. “Our model's multilingual strength and consistently high F1 scores give organizations that durable protection.”
AvailabilityThe ActiveFence AI Security Benchmark Report: Prompt Injections is available today. To learn how ActiveFence's Guardrails and Red Teaming products deploy this model in production and how ActiveFence secures seven of ten leading LLM providers visit ActiveFence.com.
About ActiveFenceActiveFence is the leading provider of AI security and safety solutions for online experiences and AI applications, safeguarding more than 3 billion users, top foundation models, and the world's largest enterprises and tech platforms every day. As a trusted ally to major technology firms and Fortune 500 brands that build user-generated and GenAI products, ActiveFence secures applications against prompt injection and other attacks with Real-Time Guardrails and continuous Red Teaming. Powered by deep threat intelligence, unmatched harmful-content detection, and coverage of 117+ languages, ActiveFence enables organizations to secure their applications, deliver engaging and trustworthy experiences at global scale while operating safely and responsibly across all threat landscapes.
https://mma.prnewswire.com/media/2753851/Comparative_Results_Infographic.jpg
https://mma.prnewswire.com/media/2753852/F1_Performance_Per_Language_and_Model_Infographic.jpg
https://mma.prnewswire.com/media/2753853/FPR_Per_Model_Infographic.jpg
https://mma.prnewswire.com/media/2753854/Model_Performance_Comparison_Infographic.jpg
https://c212.net/c/img/favicon.png?sn=PH54816&sd=2025-08-20
View original content to download multimedia:https://www.prnewswire.com/news-releases/activefence-releases-ai-security-benchmark-industry-leading-precision-and-f1-in-detecting-prompt-injections-302534029.html
SOURCE ActiveFence
https://rt.newswire.ca/rt.gif?NewsItemId=PH54816&Transmission_Id=202508201000PR_NEWS_USPR_____PH54816&DateId=20250820