Microsoft Security Chief Warns: AI Attacks Have Entered the Machine Speed Era

AI Hacking Incident Sounds Industry Alarm

In an exclusive interview with Bloomberg, Hayete Gallot, Executive Vice President of Microsoft's security business, sent a warning that should put the entire tech industry on alert: AI-driven cyberattacks are unfolding at "machine speed," and traditional security defenses are facing unprecedented challenges.

"Machine speed" has a specific technical meaning in cybersecurity. The lifecycle of traditional cyberattacks—from reconnaissance, weaponization, delivery, exploitation to lateral movement—typically takes days to weeks. But AI-driven attacks can compress this cycle to minutes or even seconds. Large language models can analyze tens of thousands of lines of code and identify potential vulnerabilities in seconds, automated tools can instantly generate targeted exploit code, and AI-orchestrated botnets can launch coordinated attacks against thousands of targets simultaneously. Each stage of the attack chain described in the MITRE ATT&CK framework is being accelerated and automated by AI one by one.

Gallot cited a highly representative case—last week's AI-driven attack involving OpenAI and Hugging Face. "They were attacked by AI and had to defend with AI," she said. "They had to adapt in real-time, build security workflows, to figure out what was actually happening in their environment." In her view, this is not an isolated incident but a microcosm of a broader trend that Microsoft observes across all its customers through its threat intelligence system.

Notably, Hugging Face, as the world's largest open-source AI model hosting platform, hosts over 500,000 models and 250,000 datasets, making it critical infrastructure in the AI supply chain. Attacks against such platforms are particularly dangerous because they can trigger a "supply chain attack" effect—if attackers can inject backdoors into popular models or tamper with model weights, all downstream enterprises and applications using those models will be affected. These attacks are similar to the 2020 SolarWinds incident, but due to the black-box nature of AI models, malicious tampering is harder to detect and the impact could be even broader.

The reason this incident is being called an "industry wake-up call" is that it reveals a new attack-defense paradigm: both attackers and defenders are using AI, and the pace of confrontation has leaped from human reaction speed to machine reaction speed. When attacks can be automated and scaled in an instant, traditional security models that rely on manual analysis and response will become unsustainable.

From "Signals" to "Protection": The Core Philosophy of Project Perception

Facing machine-speed attacks, Microsoft's answer is an agentic cybersecurity system called Project Perception. Gallot emphasized that the core philosophy of this new technology stack is: customers don't need more "signals"—they need real "protection."

The term "Agent" here is a core concept in AI, referring to AI systems that can autonomously perceive their environment, make decisions, and execute actions. Unlike traditional passive AI tools, agents possess goal-orientation, autonomous planning capabilities, and environmental adaptability. In the cybersecurity context, "agentic" means the system can not only detect threats but also autonomously decide on response strategies and execute defensive actions—such as automatically isolating infected nodes, patching vulnerabilities in real-time, or dynamically adjusting firewall rules—without waiting for human security analysts to intervene. This paradigm draws on research in reinforcement learning and multi-agent systems, evolving security systems from "tools" into "collaborators."

"Today you have a ton of signals, but signals are ultimately just signals," she explained. "Customers don't want more signals—they want protection. That's exactly what we're trying to do with AI—truly arm them so they have the ability to protect themselves."

This statement highlights a widespread pain point in enterprise security: security tools often generate massive volumes of alerts and logs, but there's an enormous gap between "detecting anomalies" and "taking action." This problem is known in the industry as "alert fatigue" and has plagued Security Operations Centers (SOCs) for years. According to research by the Ponemon Institute, enterprises face an average of over 10,000 security alerts per day, yet security teams typically can only investigate fewer than 5% of them. The high volume of false positives causes analysts to gradually become desensitized to alerts, and genuine threat signals end up buried in noise. While traditional SIEM (Security Information and Event Management) systems excel at collecting and correlating logs, they remain significantly deficient in extracting actionable intelligence from massive datasets.

The goal of agentic systems is precisely to introduce AI autonomous decision-making and execution capabilities at this juncture, transforming passive signal collection into proactive real-time protection. From Microsoft's product evolution perspective, the previously launched Security Copilot is a GPT-4-based security assistant positioned as a "copilot" to assist human analysts; Project Perception represents a further leap from "assisting human analysis" to "autonomous agent-based defense," marking the evolution of security AI from copilot mode to autopilot mode.

How AI Amplifies the Cyber Attack Surface

The most insightful part of the interview was Gallot's answer to the question of "where attacks come from." She pointed out that cybersecurity has multiple attack vectors, and a widely discussed focus recently has been software vulnerabilities.

"In the past, you could only find this many vulnerabilities," she gestured. "But now, in the blink of an eye, you can find thousands upon thousands of vulnerabilities—which means your attack surface is expanding dramatically. Attackers can also instantly assemble thousands of attack forces against you."

This reveals AI's fundamental impact on the cybersecurity landscape: it dramatically lowers the cost of vulnerability discovery and exploitation. AI's application in vulnerability discovery has moved from theory to practice. Google's Project Zero team has used AI-assisted fuzzing techniques to discover hundreds of previously unknown vulnerabilities. In 2024, AI systems demonstrated the ability to autonomously discover and patch software vulnerabilities at DARPA's AI Cyber Challenge (AIxCC). More notably, large language models (such as GPT-4, Claude, etc.) have been proven capable of analyzing source code and identifying multiple vulnerability classes including SQL injection, buffer overflow, and race conditions. This means that vulnerability mining work that previously required specialized security researchers weeks to complete can now be done by AI in bulk within extremely short timeframes—a boon for defenders and a weapon for attackers alike.

Historically, when we discussed malicious actors, we typically pointed to nation-state hacking organizations or sophisticated hacker groups—all requiring considerable resources and technical expertise. But now, AI makes vulnerability exploitation highly accessible, dramatically lowering the barrier. A novice attacker armed with AI tools may achieve attack effects that previously only senior hackers could accomplish. This "democratization of capabilities" in the security context means that the number and diversity of threat actors are increasing dramatically.

Gallot acknowledged this assessment: "With AI, we've made vulnerability exploitation much easier. Customers are witnessing this firsthand, which is why they're also embracing AI at lightning speed."

The AI Arms Race on Both Sides of Attack and Defense

The deeper implication of this conversation is that cybersecurity is evolving into an AI-versus-AI arms race. When attackers can use AI to scan code in minimal time, discover massive numbers of vulnerabilities, and automate attacks, the only viable response for defenders is to similarly use AI to discover vulnerabilities in real-time and patch them before they can be exploited.

The scenario Gallot described has become a "top concern" for every enterprise customer: when scanning their own code, the number of vulnerabilities they discover is surging due to AI's involvement. This is no longer a theoretical risk—it's happening right now.

From an industry perspective, Microsoft's strategic intent in launching Project Perception is clear: in the enterprise security market, whoever can first deliver mature agentic defense capabilities will seize the initiative in next-generation security competition. Microsoft's security business has become one of its fastest-growing segments, with annual revenue exceeding $20 billion, covering full-stack security capabilities including identity management (Entra ID), endpoint protection (Defender), cloud security (Sentinel), and compliance governance. Microsoft processes over 78 trillion security signals daily, and its threat intelligence network covers billions of devices worldwide. This massive data foundation provides a unique advantage for training security AI models, giving Microsoft a significant first-mover resource barrier in the competition for agentic security systems.

This also reflects a fundamental shift in the product logic of security vendors—from providing "observability" and "alerts" to providing "autonomous response" and "proactive protection." Major competitors like CrowdStrike, Palo Alto Networks, and SentinelOne are also accelerating their deployment of AI-native security capabilities. The entire industry stands at a historic inflection point, transitioning from Detection & Response to Autonomous Protection.

Key Takeaways for Enterprises and Practitioners

Synthesizing Gallot's perspectives, this interview conveys several highly valuable signals for enterprises and practitioners:

First, the speed dimension of competition has fundamentally changed. When attacks unfold at machine speed, the window for manual response is compressed to near-zero. Automated, intelligent defense is no longer optional—it's a necessity. Enterprises need to reassess their security operations time metrics—Mean Time to Detect (MTTD) and Mean Time to Respond (MTTR)—and recognize that in the age of AI attacks, these metrics must be compressed from hours to seconds.

Second, the expansion of the attack surface is irreversible. AI has made vulnerability discovery scalable. Enterprises need to reassess their exposure risk and incorporate AI-assisted vulnerability management into routine security processes. This includes not only continuous scanning of proprietary code but should also cover open-source dependencies, third-party API interfaces, and new attack surfaces like AI model supply chains.

Finally, the value criteria for security tools are being reshaped. The value of "signals" is declining, while the value of "protection" and "action" is rising. When selecting security solutions, enterprises should focus more on whether they possess real-time, autonomous protection and remediation capabilities, rather than merely alert volume. The evaluation dimension should shift from "how many threats can it detect" to "how many threats can it autonomously neutralize in how short a time."

It's foreseeable that as AI deeply penetrates both sides of attack and defense, the cybersecurity field will undergo a profound transformation centered on agentic technology. And the outcome of this transformation will largely depend on who can better harness "machine speed."