With an Israeli touch: Microsoft unveils a cyber defense system based on AI agents

Microsoft is unveiling this evening (Monday) a new cyber defense system based on artificial intelligence agents called Perception, which it says is designed to boost cyber defense in the AI era. At the same time, the company is presenting an AI model for cyber that it says is capable of providing better performance in identifying security vulnerabilities than Anthropic's Mythos. Some of the capabilities revealed by the company were developed at the tech giant's Israeli R&D center.

Last week, OpenAI revealed that advanced models it tested managed to escape their closed testing environment to the open internet and break into the computer systems of the Hugging Face platform, in an operation that lasted hours but would have required a skilled human days to complete. The incident clarified that the future that cyber experts warned about, in which autonomous AI agents carry out sophisticated and dangerous hacks without human intervention, is actually already a reality. Microsoft's new cyber defense system, Perception, was built, according to the company, specifically for this new reality.

"Artificial intelligence is changing the rules of the game in the cyber world," the company said in a press release. "Today, attackers can develop attacks at a higher speed, scale them in unprecedented volumes, and customize them, while organizations are required to protect more complex digital environments that include identities, applications, clouds, data, and AI systems. In this reality, traditional security solutions, built for a world where humans made most of the decisions, struggle to cope with AI-based attacks operating at unprecedented speed and scale."

At the center of Perception are AI models and autonomous agents that use relevant internal organizational information for each company to create proactive defense that focuses on identifying and mitigating cyber threats. It relies on three types of agents that work together:

• Red Team agents that identify breaches and possible attack vectors in the organization's systems

• Blue Team agents that examine the findings against the organizational context and assess which ones pose a real risk

• Green Team agents that implement the findings for the purpose of performing remediation actions and closing breaches

The agents' work leverages what Microsoft calls Security Context – an up-to-date snapshot of organizational assets, user identities, applications, cloud environments, AI systems, and relevant cyber threats. "This layer allows the AI agents to understand the meaning of each security incident, prioritize risks, and make decisions more accurately, consistently, and efficiently, while reducing the computing costs required to operate the system at a large scale. Instead of each agent having to collect and analyze information by itself, they all work based on the same shared snapshot," the company stated.

Alongside the system, Microsoft also revealed its first AI model for cyber of its own development. The model, AI-Cyber-1-Flash, is intended for identifying security vulnerabilities in software systems, similar to Anthropic's Mythos model. According to Microsoft, when the model works in combination with MDASH (an AI agent-based security scanning system that the company revealed in May), it receives a score of 96 in the CyberGym test, 12 points higher than Mythos.

A significant part of the new capabilities is based on work carried out at Microsoft's R&D center in Israel. According to the company, the system was developed in Israel and will be the first solution to implement Microsoft's cyber model. "Thus, one of the first capabilities that realizes in practice the vision of Project Perception is based on Israeli development, as part of the new generation of defense solutions that Microsoft is developing for the artificial intelligence era," it was stated.