AI cybersecurity leader CyCraft announced a strategic partnership with APMIC, a provider of model-as-a-service solutions, to launch the industry's first AI model designed specifically for defense: XecGuard. Paired with a model security evaluation service, this initiative leverages a 3-billion (3B) parameter open-source model developed by APMIC and the Twinkle AI community.
Both CyCraft Technology and APMIC were founded in 2017. CyCraft recently revealed plans to apply for listing on Taiwan's Innovation Board. Meanwhile, APMIC is a member of NVIDIA's startup program and has attended the Nvidia GTC conference in the US.
According to CyCraft, this collaboration aims to address the security challenges enterprises face when adopting AI by introducing a top-tier protection layer to defend against malicious attacks, including prompt injection, prompt extraction, etc.
XecGuard is an innovative LoRA-based security upgrade module built upon nearly 60,000 proprietary security offense-defense datasets and a multi-stage fine-tuning technique. It can be directly installed on any large language model (LLM) to provide powerful defensive capabilities without modifying the model architecture.
CyCraft conducted real-world tests using the 3B open-source model in combination with the XecGuard security module. The system underwent red team adversarial testing, and results showed an average security improvement of 19.4%, with up to 33.9% improvement in specific attack scenarios, which demonstrates strong resilience against threats.
Additionally, the team highlighted that F1 equipped with XecGuard outperformed commercial LLMs of similar or larger sizes, such as Llama 3B, Qwen3 4B, Gemma3 4B, and DeepSeek 8B. It is particularly strong in resisting prompt injection and prompt extraction attacks, all while maintaining high response quality. The findings also prove that through secure distillation and dedicated training, smaller models can rival international commercial-grade large models.
Benson Wu, CyCraft CEO and co-founder, stated that in the past, AI was used to solve information security problems; now, AI is used to protect AI. XecGuard incorporates real-world defense experience across critical sectors like government and finance, offering enterprises the most robust protection in the AI era.
Jerry Wu, APMIC founder and CEO, agreed and emphasized that security must be a default feature, not an optional one. Only when built on a secure foundation will enterprises embrace AI without hesitation.
Article edited by Jack Wu
