Hackers Accelerate Attacks with AI? Microsoft Launches Project Perception, Building an Automated Vulnerability Patching 'Green Team' Defense Network

AI is not only accelerating enterprise development and operations but also lowering the barriers to cyberattacks. As hackers begin leveraging AI and autonomous agents to identify vulnerabilities at lower cost, higher speed, and larger scale, traditional security architectures—still heavily reliant on manual investigation and incident response—may struggle to keep pace with the 'machine-speed' rhythm of modern attack and defense. Microsoft (Microsoft, NASDAQ: MSFT) today (28th) announced the launch of its new agentic cybersecurity system, Project Perception, and unveiled MAI-Cyber-Flash-1, its first security-specific AI model designed for software vulnerability analysis, aiming to fundamentally reshape enterprise cybersecurity defenses.

Unlike past defense models centered on massive detection, alerts, and post-incident handling, Project Perception integrates security signals, environmental context, AI models, and specialized agents into a continuously operating closed-loop system. It not only helps enterprises identify risks but also analyzes attack paths, prioritizes incidents, and further transforms insights into remediation actions.

Microsoft states clearly that the key to next-generation security systems is no longer about generating more alerts, but whether they can "continuously perceive, reason, and act." Project Perception is scheduled to enter public preview on August 3.

AI Lowers Attack Costs; Traditional Defenses Lag in Speed

As enterprises adopt more identities, endpoints, applications, data, cloud, and AI systems, the digital environments requiring protection become increasingly fragmented and complex. Meanwhile, attackers can use AI to automate vulnerability discovery, social engineering, and attack processes, further reducing the time and cost required to launch attacks.

Microsoft points out that past security architectures were primarily designed for human attackers. However, as AI, agents, and machine-speed attacks become more prevalent, defenders still need to manually integrate signals from vast data, reconstruct scenarios, and assess risks—making it difficult for security teams to respond timely to evolving threats.

To address this, Microsoft proposes a new Cyber Stack concept, requiring security systems to continuously monitor the entire digital environment, perform reasoning across massive contexts, take action at near-machine speed, and continuously learn as enterprise environments and threats evolve. However, Microsoft emphasizes that automation does not mean fully handing control to AI. Project Perception is designed to assist defenders in enhancing judgment and action capabilities, with final control remaining in human hands.

Microsoft believes that achieving agentic security requires not just adding agents to existing workflows, but a completely new Cyber Stack designed from the ground up. (Provided by Microsoft)

Three Types of Agents—Red, Blue, Green—Collaborate to Build a Closed-Loop Defense

The core of Project Perception is coordinating three types of specialized agents with distinct roles, continuously examining enterprise security from attack, investigation, and remediation perspectives.

Red team agents act as attackers, proactively identifying potential attack paths that could lead to system compromise before they are exploited. Blue team agents are responsible for investigating security incidents, combining enterprise environment and threat context to reason which events pose real risks. Green team agents take corrective actions, improving configurations, reducing exposure, and strengthening overall defense.

These three agent types do not operate independently but form a continuous defensive cycle. After red team agents identify potential attack paths, blue team agents assess risk and priority, then green team agents execute improvements. The system then re-evaluates the results, shifting security defense from post-incident response to continuous risk discovery, assessment, and mitigation.

Microsoft states that Project Perception’s operational capability is built upon four foundations: visibility into the digital environment, actionable security actions, long-term accumulated threat intelligence and defense experience, and AI model capabilities. The system covers identity, endpoints, applications, data, cloud, and AI systems, converting analysis results into real actions via Microsoft Security products.

Security Context Connects Environmental Information—No Need to Reconstruct Risk Each Time

However, simply adding more agents is insufficient to create an autonomous security system. The key to accurate risk assessment lies in whether different agents can access consistent, real-time, and complete environmental information.

Therefore, Microsoft introduces "Security Context" in Project Perception, linking together fragmented security data, knowledge, and semantics across the enterprise digital environment. This continuously builds a comprehensive view of assets, identities, relationships, risks, and activity states, enabling different agents to share near real-time organizational context.

In traditional approaches, security analysis tools or AI models often need to repeatedly gather information from raw signals, analyze relationships between assets, and reconstruct event context. Project Perception allows agents to directly access required information from the continuously updated Security Context, accelerating reasoning and decision-making while reducing the tokens, computing resources, and costs required for large-scale operations.

In other words, the difference lies not just in whether AI is used, but in whether different AI agents can collaborate based on a shared understanding of the enterprise environment, avoiding judgment gaps caused by individually interpreting fragmented data.

Avoiding Reliance on a Single Large Model—Dynamically Selecting the Best Tool by Task

Since enterprise security must operate continuously, model accuracy is not the only consideration—reliability, latency, and usage cost also affect large-scale deployment.

Microsoft believes no single AI model can perform optimally across all security tasks. Therefore, Project Perception adopts a multi-model architecture, dynamically selecting between frontier AI models (Frontier Models) and security-specialized models (Specialized Models) based on task requirements, balancing quality and cost.

For example, tasks requiring understanding of complex attack contexts may rely on high-reasoning frontier models, while specific vulnerability analysis or highly repetitive security tasks can be handled by faster, lower-cost specialized models trained on domain-specific data.

Microsoft’s security research team will continuously evaluate which models are best suited for different tasks through real-world security scenarios, benchmarks, and assessments—ensuring enterprises are not locked into a single model. This reflects a shift in competitive focus after adopting agentic AI—from merely pursuing model capability to managing diverse models, agents, and workflows.

MAI-Cyber-Flash-1 Targets Vulnerability Analysis First—Achieves 96% in CyberGym Evaluation

Aligned with Project Perception’s multi-model strategy, Microsoft also launched its first in-house developed security-specific model, MAI-Cyber-Flash-1, initially targeting software vulnerability management and analysis.

Microsoft has integrated this model into its multi-model agent group for software vulnerabilities, MDASH. According to official results, MDASH with MAI-Cyber-Flash-1 achieved 96% effectiveness in the CyberGym security evaluation—12 percentage points higher than Mythos. Under the same conditions as the current MDASH configuration, it also reduces costs by nearly half.

These results indicate that, with abundant historical vulnerability data and specialized security training content, a fine-tuned specialized model does not need to fully rely on larger general-purpose models to achieve superior performance and economic efficiency in specific tasks.

Microsoft states that software vulnerability management is just the first application scenario for MAI-Cyber-Flash-1, and Project Perception will further integrate this model into other security workflows in the future.

From Detection and Alerts to Autonomous Action—Microsoft Redefines the Cyber Stack

To support agentic security, Microsoft divides its new Cyber Stack into multiple interconnected layers. At the base, signals and sensors monitor the enterprise environment. Security Context converts raw data into content understandable by agents. Models provide intelligence and reasoning. Harness orchestrates models, agents, and workflows. Agents execute security tasks, and finally, actuators convert decisions into actual protective measures.

For example, when the system identifies that an identity permission, endpoint vulnerability, and cloud configuration could be chained into an attack path, it does not just issue an alert. It evaluates the risk level, determines the handling order, and then takes actions—such as reducing permissions, modifying configurations, or patching vulnerabilities—through existing security products.

This signifies a shift in enterprise security architecture—from the past model centered on point products and manual analysis—toward shared context, multi-model collaboration, and autonomous AI-driven defense.