Softfreak Registers 2 Patents for AI Security Technology, Expands Vulnerability Testing and Access Control

As corporate adoption of AI has expanded and cyberthreats have become more sophisticated, the security industry’s paradigm is shifting. The industry is focusing on AI control and management for safe AI use, while also concentrating on applying AI in the security field to respond to AI attacks.

In light of this trend, Softfreak has continued to research and file patents for AI-related technologies. Softfreak announced on the 12th that 2 patents for AI security technology have been registered.

The 2 registered patents cover AI privacy vulnerability diagnosis technology and AI gateway security control technology. Among them, the patent for 'Multimodal Attack Method for Diagnosing Privacy Vulnerabilities in Artificial Intelligence Models' is one of 5 AI privacy vulnerability diagnosis patents Softfreak filed last year.

Multimodal AI processes text and images together. However, when checking vulnerabilities in multimodal AI, text-based attacks alone make it difficult to verify enough.

The technology diagnoses in advance the possibility of sensitive information leakage from AI models through attack scenarios that use multiple input types. The technology is intended to verify the safety of the AI model itself, and it expands the scope of security into 'AI Security Testing.'

Softfreak’s patent is titled 'Security Control Method for Prompts and System Behaviors Executed by an AI Gateway.' The patent is a technology that controls prompts and system behaviors between AI and enterprise systems based on security policies. It responds to the changing environment in which AI agents act as a new 'system user' and 'API Consumer,' allowing enterprises to control prompts delivered by AI, control access permissions, and control and track the use of data and systems.

Softfreak has accumulated expertise in API management and security, and it has authentication, authorization, access control, and policy enforcement technologies. Based on this, it is developing next-generation gateway technology that connects API gateways to LLM gateways and then to AI gateways.

Softfreak is also researching security automation technology using AI. It has also filed a patent for technology that analyzes LLM-based security guidelines and then automatically generates vulnerability inspection rules and diagnostic scripts. In addition, it is pursuing integration with the vulnerability diagnosis solution F-Avata.

Softfreak continues to file patent applications related to technologies that protect generative AI models. The specific technologies it has continued to apply for include 'risk-based adaptive policy orchestration,' a protection policy adjustment technology that tunes protection policies by risk level; 'learning transfer perturbation,' a protection technology that strengthens resistance to retraining; and 'learning-transfer watermarking,' a protection technology that tracks retraining, distillation, and fine-tuning of downstream models using model outputs.

Softfreak plans to combine its existing product lineup with AI security technology. Its existing products include API gateways, vulnerability diagnosis solutions, and account management and access control solutions. Based on these, it plans to establish an 'AI Security & API Governance' framework. The framework will cover security task automation, AI model safety verification, AI agent access and behavior control, and protection of generative AI models.

Explaining the background to these plans, Kim Dong-cheol, CEO of Softfreak, said that attackers are using AI to automate vulnerability discovery and attacks, making it difficult for human-centered responses alone to keep up with the speed and scale of attacks. He added that defenders also need to actively use AI, and that enterprises need to verify the safety of the AI they use and control access and behavior.

Source: TECHWORLD · Kim Hye-jin
Original: https://www.epnc.co.kr/news/articleView.html?idxno=408035

References

• https://www.epnc.co.kr/news/articleView.html?idxno=408035