Microsoft unveils agentic AI stack to streamline security operations

Key Takeaways

• MDASH is a new Microsoft security initiative built by the internal FORGE team, which combines offensive research with generative AI techniques.
• The FORGE team is led by Georgia Tech professor Taesoo Kim and includes many of his current and former PhD students, several of whom were part of Team Atlanta, the winners of the DARPA AI Cyber Challenge (AIxCC).
• Project Perception creates a dynamic security graph by fusing Microsoft’s threat intelligence, security telemetry, and detailed knowledge of customer environments.
• The graph powers a trio of specialized agents—red, blue, and green teams—that autonomously execute vetted playbooks to investigate, defend, and remediate threats.
• Human analysts remain in the loop: they can task Perception with high‑level goals (e.g., “Are we protected against this new threat actor?”) and the system automatically distributes sub‑tasks to the appropriate agents.
• A live demonstration showed Perception ingesting a fresh threat‑intel report, correlating IOCs and TTPs against the security graph, and launching coordinated red‑team probing, blue‑team detection tuning, and green‑team mitigation actions without manual scripting.
• The approach aims to shrink mean‑time‑to‑detect (MTTD) and mean‑time‑to‑respond (MTTR) by augmenting SOC staff with AI‑driven, play‑book‑guided autonomy while preserving expert oversight.
• Challenges include managing false‑positive rates, ensuring data privacy across telemetry sources, and maintaining rigorous validation of AI‑generated playbooks.
• Microsoft plans to expand Perception’s agent library, deepen integration with Azure Sentinel, Defender, and Purview, and publish research findings from the FORGE team to the broader security community.
• Ultimately, MDASH and Project Perception represent a shift toward self‑orchestrating, knowledge‑driven security operations that blend human expertise with scalable AI agents.

Introduction to MDASH and Its Strategic Vision

MDASH (Microsoft Defensive Autonomous Security Hub) was unveiled in May as the latest addition to Microsoft’s growing portfolio of AI‑enhanced security solutions. Rather than a single product, MDASH is an overarching framework that couples offensive research concepts with generative artificial intelligence to produce systems capable of proactive threat hunting, autonomous response, and continuous security posture assessment. The initiative reflects Microsoft’s belief that the future of cybersecurity lies in blending human analyst intuition with machine speed and scale, thereby enabling organizations to stay ahead of increasingly sophisticated adversaries.

The FORGE Team: Origins, Leadership, and Pedigree

MDASH is the brainchild of a newly formed internal Microsoft unit called FORGE (Frontier Offensive Research & Generative Exploration). FORGE is led by Taesoo Kim, a professor of computer science at the Georgia Institute of Technology renowned for his work in systems security, binary analysis, and automated vulnerability discovery. The team draws heavily from Kim’s research group, comprising many of his current and former PhD students who bring deep expertise in offensive techniques, program synthesis, and machine‑learning‑driven reasoning. Notably, several members of FORGE were core contributors to Team Atlanta, the victorious squad in the DARPA AI Cyber Challenge (AIxCC), a two‑year competition that tasked participants with building autonomous AI systems capable of detecting, patching, and defending critical open‑source software. This pedigree equips FORGE with a rare combination of offensive ingenuity and defensive rigor, forming the technical foundation for MDASH’s autonomous capabilities.

Project Perception: Constructing the Living Security Graph

At the heart of MDASH lies Project Perception, a platform designed to synthesize disparate data streams into a unified, continuously updated security graph. Perception ingests three primary sources:

1. Threat Intelligence Feeds – curated reports, IOCs (indicators of compromise), and TTPs (tactics, techniques, and procedures) from Microsoft’s global threat intelligence operation and partner feeds.
2. Security Telemetry – real‑time event logs, network flows, endpoint behaviors, and cloud activity harvested from Azure Sentinel, Microsoft Defender for Endpoint, Identity, and Cloud Apps.
3. Environmental Knowledge – detailed models of customer assets, configurations, patch levels, user roles, and business‑critical workloads derived from configuration management databases (CMDB), asset inventories, and organizational policies.

By correlating these inputs, Perception builds a dynamic graph where nodes represent assets, users, processes, and vulnerabilities, while edges capture relationships such as data flow, trust boundaries, and exploit pathways. This graph is refreshed in near‑real time, allowing the system to reflect the latest threat landscape and infrastructure changes instantly.

Specialized Agents: Red, Blue, and Green Teams in Action

Perception’s graph serves as the operational playground for three categories of specialized agents, each analogous to a traditional security team but implemented as autonomous AI modules:

• Red Team Agents – emulate adversary behavior. Given a target objective (e.g., “gain privilege escalation on a Windows server”), these agents simulate attack paths using the graph to identify feasible exploit chains, generate proof‑of‑concept payloads, and test defenses in a safe, contained sandbox.
• Blue Team Agents – focus on detection and mitigation. They continuously monitor the graph for anomalous patterns, update detection rules, prioritize alerts based on potential impact, and recommend configuration hardening or patching actions.
• Green Team Agents – handle orchestration, remediation, and recovery. Once a threat is confirmed, green agents initiate containment steps (e.g., isolating a compromised host), trigger automated patch deployment, and coordinate recovery workflows such as forensic data collection and post‑incident reporting.

Each agent operates according to a library of playbooks—pre‑validated, step‑by‑step procedures encoded as executable policies. Playbooks are vetted by human experts, continuously refined through feedback loops, and can be invoked independently or chained together to accomplish complex missions.

Human‑In‑the‑Loop Workflow: Analyst Tasking and Agent Dispatch

Although the agents possess considerable autonomy, Microsoft designed by‑level autonomy, the system maintains a human‑in‑the‑loop model to ensure accountability and contextual nuance. A security analyst interacts with Perception through a natural‑language interface or a structured query language. For example, upon receiving a fresh threat‑intel bulletin about a novel threat actor, the analyst might issue the command:

“Investigate whether our environment is protected against the threat actor XYZ and identify any gaps.”

Perception parses the request, extracts the relevant IOCs and TTPs from the bulletin, and maps them onto the security graph. It then determines which combination of red, blue, and green agents is best suited to answer the analyst’s question, dispatches the appropriate playbooks, and begins executing tasks in parallel. Throughout the operation, the analyst receives real‑time status updates, can intervene to adjust priorities, and ultimately reviews a consolidated report that outlines findings, recommended actions, and any residual risk.

Demonstration Scenario: From Threat Report to Autonomous Response

During MDASH’s launch event, Microsoft showcased a concrete illustration of this workflow. A simulated threat‑intel report arrived describing a new ransomware group that leverages a recently disclosed zero‑day vulnerability in a widely used library, employing spear‑phishing emails with malicious Office documents as the initial vector. The analyst tasked Perception with assessing the organization’s exposure.

1. Graph Enrichment – Perception ingested the report’s IOCs (specific file hashes, C2 domains) and TTPs (phishing with macro‑laden docs, exploitation of CVE‑2024‑XXXX, lateral movement via SMB). These were instantly linked to corresponding nodes in the security graph (e.g., vulnerable library instances, email gateway configurations, SMB share permissions).
2. Red Team Probe – Red agents simulated the phishing delivery path, attempting to send a benign macro‑enabled document to internal mailboxes and checking whether email security gateways blocked it. Simultaneously, they probed internal hosts for the vulnerable library version, attempting to exploit CVE‑2024‑XXXX in a controlled container.
3. Blue Team Detection Tuning – Blue agents reviewed alert streams from Defender for Office 365 and Azure Sentinel, identifying any missed detections. They automatically generated updated hunting queries and refined email attachment sandbox rules to catch the macro technique.
4. Green Team Remediation – Upon confirming that a subset of workstations still ran the vulnerable library, green agents orchestrated a targeted patch rollout via Windows Update for Business, isolated the affected machines from the network, and initiated forensic logging for further analysis.
5. Outcome Synthesis – Perception compiled a concise briefing: the organization’s email gateway blocked 98% of simulated phishing attempts, but two legacy servers remained unpatched, presenting a residual risk. The report recommended immediate patching and a review of legacy system management procedures.

The entire sequence unfolded with minimal manual scripting, highlighting how Perception can close the loop between intelligence ingestion and defensive action within minutes rather than hours or days.

Technical Underpinnings: AI Models, Reasoning Engines, and Integration

The autonomy demonstrated by Perception rests on several AI‑enhanced components:

• Large Language Models (LLMs) fine‑tuned on security‑specific corpora enable the system to understand natural‑language threat reports, extract structured IOC/TTP data, and generate human‑readable summaries.
• Graph Neural Networks (GNNs) propagate information across the security graph, scoring nodes and edges for risk likelihood based on topological features and temporal telemetry trends.
• Reinforcement Learning (RL) policies train red and blue agents to optimize attack‑defense sequences, balancing exploration (novel TTP discovery) with exploitation (high‑yield detection or mitigation).
• Symbolic Reasoning Engines ensure that playbook execution adheres to organizational policies, regulatory constraints, and safety guards (e.g., never executing destructive actions on production systems without explicit approval).

These modules are hosted on Azure Kubernetes Service (AKS) clusters, with tight integration to Microsoft’s existing security stack—Sentinel for SIEM, Defender for endpoint/cloud/identity, and Purview for data governance—allowing Perception to read telemetry, write detections, and trigger automated responses via native APIs.

Potential Impact and Operational Benefits

Deploying MDASH and Project Perception promises several tangible advantages for SOCs:

• Reduced MTTD/MTTR – By automating the correlation of fresh threat intelligence with internal asset maps, analysts spend less time on manual data gathering and more on strategic decision‑making.
• Scalable Threat Hunting – Red‑team agents can continuously probe the environment for emerging attack patterns, effectively providing an always‑on adversarial emulation layer.
• Consistent Playbook Execution – Encoded playbooks eliminate variability in response quality, ensuring that best‑practice actions are applied uniformly across incidents.
• Enhanced Situational Awareness – The live security graph offers a visual, up‑to‑date view of risk exposure, facilitating executive reporting and risk‑based budgeting.
• Optimized Analyst Utilization – Routine, repeatable tasks are handled by agents, freeing senior analysts to focus on complex investigations, threat‑intel authoring, and proactive security strategy.

Challenges, Limitations, and Mitigation Strategies

Despite its promise, the approach introduces considerations that must be addressed:

• False Positives/Negatives – Over‑reliance on AI‑generated detections can lead to alert fatigue or missed threats. Mitigation includes confidence scoring, human validation loops, and continuous model retraining on labeled data.
• Data Privacy and Sovereignty – Aggregating extensive telemetry raises concerns about data leakage. Microsoft enforces strict access controls, encryption, and regional data residency options to satisfy compliance regimes (GDPR, CCPA, etc.).
• Explainability – Stakeholders need to understand why an agent flagged a particular asset or recommended a specific action. Perception incorporates explainable AI techniques (e.g., attention visualizations in GNNs, rule‑trace logs) to provide transparent rationales.
• Safety of Autonomous Actions – Unintended disruption from automated remediation is a risk. Green‑team agents are gated by approval workflows for high‑impact actions (e.g., network isolation, credential reset) and operate in simulation mode first for validation.
• Model Drift and Adversarial Evasion – As threat actors adapt, AI models may degrade. Ongoing red‑team exercises and adversarial training regimens are built into the FORGE research agenda to keep models robust.

Future Roadmap and Vision

Microsoft envisions MDASH evolving into a platform where customers can compose their own AI‑driven security agents using low‑code tooling, drawing from a shared repository of playbooks contributed by Microsoft, partners, and the community. Near‑term priorities include:

• Expanding the agent library beyond red/blue/green to include purple team (collaborative attack‑defense simulation) and gold team (strategic risk & compliance) agents.
• Deepening integration with Microsoft Security Copilot to allow natural‑language orchestration of complex, multi‑step investigations.
• Publishing research outputs from the FORGE team—particularly novel techniques in generative exploit synthesis and graph‑based threat forecasting—to advance the broader security field.
• Offering industry‑specific extensions (e.g., healthcare, finance) that tailor the security graph and playbooks to sector‑specific regulations and threat profiles.

Through these developments, MDASH aspires to move security operations from a reactive, analyst‑intensive posture toward a self‑optimizing, knowledge‑driven ecosystem where human expertise and AI agents collaborate seamlessly to protect critical assets at machine speed.