VirtueGuard: Enterprise-Grade AI Security and Safety Now on Together AI
Together AI
Together AI just added VirtueGuard, a content-safety filter that checks AI outputs in 8 milliseconds. It's 50x faster than rivals like AWS or Azure, and you turn it on with one line of code.
Based on reporting by Together AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Guardrails for AI have always had an annoying tradeoff: catch the bad stuff, or keep the app fast. Together AI's new partner, Virtue AI, claims to have broken that tradeoff with VirtueGuard, a safety-check layer now live across Together's model catalog.
The headline number is 8 milliseconds. That's how long VirtueGuard takes to scan a prompt or response for trouble, compared to 400ms or more from cloud incumbents like AWS Bedrock and Azure. Together says accuracy doesn't suffer for the speed boost, either — VirtueGuard posts an 89% F1 score against 76% for Bedrock and 81% for Azure, and its false-positive rate sits at 0.058 versus AWS's 0.235. Fewer false alarms means fewer legitimate requests getting blocked, which is the kind of detail that actually matters once you're running this at scale instead of in a demo.
The team behind it has real pedigree. Bo Li, Dawn Song, Carlos Guestrin, and Sanmi Koyejo — researchers whose work underpins a good chunk of modern AI safety literature — built Virtue AI, and their customer list already includes Uber, Anthropic, NVIDIA, and Glean. That's not a startup pitching vaporware; it's infrastructure that's already been stress-tested by companies whose entire business depends on not shipping a phishing email generator by accident.
Integration is the part Together seems proudest of. You add one parameter — safety_model=
My take — AI-written commentary, not fact-checked reporting
Virtue-AI/VirtueGuard-Text-Lite
Read more about this at: Together AI