Skip to main content

Ant Group Open-Sources Two Security Models to Tame AI Agents

Ant Group's AI Security Lab dropped two open-source security models on July 13, targeting the wild west of autonomous AI agents and multimodal interactions. The models, SingGuard-NSFA and SingGuard, are designed to tackle security headaches that come when AI stops just chatting and starts doing things.

The New Frontier: AI That Acts

Remember when AI was just a fancy chatbot? Those days are fading fast. Today's intelligent agents can call tools, run code, and plan multi-step tasks on their own. But with great power comes great security risks. We've seen prompt injection attacks, permission abuse, and data leaks—like Amazon Q getting poisoned, Microsoft Copilot leaking data, and the open-source agent OpenClaw showing serious vulnerabilities. In December 2025, OWASP even released a "Top 10 Risks for Intelligent Agent Application Security." And in May 2026, China's top internet regulators jointly issued guidelines for agent security governance. The message is clear: as agents get smarter, we need smarter security.

SingGuard-NSFA: A Real-Time Brake for Agents

Think of SingGuard-NSFA as a security guard that checks every move an agent makes before it acts. It intercepts risky requests and backs up responses, building a behavioral firewall. The model is built on the CIA triad (Confidentiality, Integrity, Availability) and OWASP guidelines, categorizing risks into 7 major types, 28 subcategories, and 185 specific scenarios. It's trained on nearly 100,000 samples covering 133 languages.

What's cool? It works in two modes: one generates detailed risk reports for audits, the other zips through single-risk checks in about 50 milliseconds—perfect for high-traffic online systems. And it comes in four sizes (0.8B, 2B, 4B, 9B) to fit different setups. Tests show the tiny 0.8B model performs like an 8B model. Plus, adding new risk categories only requires training a lightweight module, not the whole model.

Image

SingGuard: The Multimodal Gatekeeper

Meanwhile, content security in multimodal AI is a whole other beast. Just last month, Anthropic's Claude Fable5 was tricked by researchers using Unicode and Cyrillic letters to bypass filters—the model understood the malicious intent, but the classifier didn't. That's the problem: as models get better at understanding distorted text and images, old keyword-based filters fall short.

SingGuard steps in with a unified security framework for text, images, and cross-modal content. It spots complex attacks hidden across different modalities. And it supports dynamic rule updates—no retraining needed. Its "fast and slow" reasoning quickly handles simple cases and digs deeper for tricky ones, balancing speed and accuracy.

Image

In head-to-head tests across 35 datasets, SingGuard beat industry leaders like Llama Guard3, Google ShieldGemma, GPT-5.1, and Gemini3-Pro on average F1 scores.

Why This Matters

Huna Ying from the China Academy of Information and Communications Technology notes that as AI shifts from content generation to autonomous execution, security must extend from content review to behavioral control and system governance. Ant Group has been at this for a while—they audited the open-source agent framework OpenClaw and co-launched a security plugin with Tsinghua University. These new models build on over two decades of Ant's security experience, from payment security to privacy protection. They're already being used in products like Ant Afu and Alipay's AI assistant.

Ant Group is also pushing standards, contributing to IIFAA's terminal agent trust specifications and leading an ITU international standard. The goal? Turn security tech into real-world practice.

Key Points

  • Ant Group open-sources two security models: SingGuard-NSFA for agent behavior monitoring and SingGuard for multimodal content filtering.
  • SingGuard-NSFA offers real-time risk detection in ~50ms, with four model sizes and easy updates.
  • SingGuard beats top competitors like Llama Guard3 and GPT-5.1 on multimodal safety benchmarks.
  • Both models address growing security threats as AI agents become more autonomous.
  • Ant Group leverages 20+ years of security expertise and contributes to international standards.