Microsoft has introduced its first artificial intelligence model built specifically for cybersecurity, pairing it with a new agentic defence system designed to help organisations find, assess and repair software weaknesses at machine speed.
The model, called MAI-Cyber-1-Flash, has been integrated into Microsoft’s MDASH vulnerability platform. It was announced alongside Project Perception, a broader security system that coordinates specialised AI agents across offensive testing, threat investigation and remediation. Microsoft says the combined approach is intended to counter attackers who are already using AI to scan code, automate reconnaissance and accelerate exploitation.
“The physics of cybersecurity are changing,” Hayete Gallot, executive vice-president of Microsoft Security, said in the company’s announcement. She argued that traditional systems built around periodic scanning and human-led response cannot keep pace with autonomous threats operating continuously.
A Cyber Model Built for Code
MAI-Cyber-1-Flash is a compact, code-heavy model derived from Microsoft’s MAI-Thinking-1 family. Unlike a general-purpose chatbot, it has been tuned for vulnerability research, including analysing large codebases, identifying suspicious paths and supporting the creation of evidence that a flaw can be exploited.
The model works inside MDASH rather than as a standalone assistant. MDASH is Microsoft’s multi-model vulnerability identification and remediation harness, built around more than 100 specialised agents. Those agents divide a security review into separate stages, including mapping the attack surface, scanning code, debating whether a suspected flaw is reachable, removing duplicate findings, generating proof-of-concept inputs and helping prepare fixes.
Gallot said the rapid growth of offensive AI leaves defenders with little room to stand still. “It’s not like we have a choice,” she said, arguing that security teams must evaluate and advance the same class of capabilities already available to attackers.
Smaller Model Handles Most Tasks
The central design choice is model routing. Microsoft says MAI-Cyber-1-Flash can handle up to 90% of MDASH tasks, while the remaining 10% of exceptionally difficult work is passed to the larger GPT-5.4 model. That arrangement is meant to avoid using an expensive frontier model for every stage of a large-scale software scan.
Microsoft reported a 95.95% score for the MAI-Cyber-1-Flash and GPT-5.4 combination on CyberGym, a benchmark based on real-world vulnerability reproduction tasks. The company rounded the result to 96% in its announcement and said it was about 12 percentage points ahead of Anthropic’s Mythos system. Rival configurations shown by Microsoft scored between 83.2% and 85.6%.
The company also claims the new configuration cuts costs by almost 50% compared with its previous leading MDASH setup, which relied on a combination of GPT-5.4, GPT-5.4 Mini and GPT-5.3 Codex. Microsoft has not published a customer price for individual scans, so the savings figure remains a comparison with its own earlier model mix rather than an independently verified market-wide cost assessment.
Project Perception Adds Agent Teams
Project Perception expands the same agentic approach beyond code scanning. It coordinates three groups of agents modelled on familiar cybersecurity roles.
Red agents examine systems from an attacker’s perspective, looking for possible paths to compromise. Blue agents investigate alerts, combine technical context and decide which findings present meaningful risk. Green agents handle corrective work, including strengthening configurations and supporting remediation.
The teams share context rather than operating as isolated tools. Microsoft says that allows a red-team finding to move directly into blue-team investigation and then into green-team repair without repeated manual handovers. The platform draws signals across identities, endpoints, applications, cloud services, data and AI systems, creating a near real-time picture of an organisation’s environment.
David Weston, Microsoft’s corporate vice-president of AI security, said the company does not intend to leave the productivity gains of AI solely to cybercriminals. “We’re not going to let the attackers have all the productivity increase,” he said.
Early Testing Shows Faster Analysis
Microsoft has already used MDASH in its own security engineering work. In May, the company said the system helped researchers discover 16 vulnerabilities across Windows networking and authentication components, including four critical remote-code-execution flaws. It also found all 21 deliberately planted bugs in a private test driver with no false positives during that run.
Nationwide Building Society is among the early organisations testing Project Perception. David Boda, the company’s chief security and resilience officer, said one threat-intelligence task was reduced from four weeks to four hours. The example suggests that the immediate value may be less about replacing security analysts and more about compressing research, triage and coordination work that normally moves between several specialist teams.
CyberGym results should still be read with care. The benchmark measures whether systems can reproduce known vulnerabilities and generate working proof-of-concept exploits. It is a useful technical comparison, but it does not on its own prove that an agent will discover every unknown flaw, produce a safe patch or make the correct decision in a live enterprise environment.
Human Approval Remains Required
Microsoft is presenting Project Perception as an agentic system, but not as an unrestricted autonomous defender. Human operators set objectives and guardrails, while high-impact actions require approval. The platform also includes role-based access controls, tenant isolation, encryption, audit logs and sandboxed execution environments without internet access.
Weston said Microsoft still has to “earn the right” to give the agents greater autonomy. That caution reflects a wider concern across the security sector: the same systems capable of finding vulnerabilities and writing patches can cause serious damage when given excessive permissions, weak identity controls or incomplete context.
Preview Begins in August
Project Perception is scheduled to enter public preview on August 3, 2026. Microsoft plans to make MAI-Cyber-1-Flash available through Azure AI Foundry using existing customer-vetting and computing-allocation processes, rather than releasing it as an unrestricted public model. The platform will initially connect its coordinated agents with Microsoft Defender, with broader expansion across Microsoft’s security products planned later.
The launch places Microsoft in a growing contest to build specialised cybersecurity models as Google, Cisco, Anthropic and OpenAI develop their own systems. Microsoft’s differentiator is its combination of an in-house model, a multi-model harness and security data drawn from more than 100 trillion daily signals and 1.6 million customers.
For security teams, the important test will not be whether the agents can produce more findings. It will be whether they consistently identify the right vulnerabilities, explain their reasoning, generate safe fixes and reduce the time between discovery and protection without creating new operational risk.
Comments