AprielGuard: ServiceNow AI Releases Advanced 8B Safeguard for Secure Agentic Workflows

ServiceNow AI has introduced AprielGuard, a high-performance safeguard model specifically engineered to secure autonomous agentic systems. Released on December 23, 2025, the 8-billion parameter model addresses the growing need for safety in multi-step AI workflows. It marks a shift from simple text moderation to a wide-ranging defense against complex, logic-based threats. Developers can access the model and related research through Hugging Face. The model identifies 16 distinct categories of safety risks, including toxicity, misinformation, and illegal activities. Beyond standard safety, it provides a dedicated defense against adversarial tactics like prompt injection and multi-turn jailbreaks. This dual-focus ensures that enterprise AI applications remain both compliant and resilient against malicious manipulation. More technical details are available in the official research paper.

Unlike traditional classifiers, AprielGuard is optimized for modern agentic ecosystems where models interact with external tools and memory. It can detect specialized attacks such as memory poisoning and tool hijacking within complex reasoning traces. This capability allows developers to monitor the internal decision-making processes of their AI agents in real time, preventing exploitation before it reaches critical systems.

Operators can choose between two primary modes of execution depending on their performance requirements. The Reasoning mode generates structured explanations for its decisions, providing valuable transparency for debugging and auditing. For production environments where speed is essential, the Fast mode offers low-latency classification without the explanatory narrative.

Built on the Apriel-1.5 architecture, the model supports a significant context window of up to 32,000 tokens. This allows it to scan extensive conversation histories and long retrieval-augmented generation documents for subtle, embedded risks. It has demonstrated strong performance across various public benchmarks, maintaining high precision in multilingual and long-form scenarios.

AprielGuard is now available for the research and developer community on the Hugging Face model hub. Its release includes detailed documentation on the synthetic data generation processes and evaluation metrics used during development. This tool represents a foundational step toward building more trustworthy and autonomous enterprise AI systems.