When Attacks Move at Machine Speed, Defense Must Too: The Case for AI-Powered Incident Response

Thousands of requests per second.

That's the operational tempo of the first documented large-scale AI-orchestrated cyber espionage campaign. Not hundreds. Not dozens. Thousands. Per second.

Anthropic's recent disclosure about disrupting this campaign revealed something security teams have feared but few have articulated: we've crossed a threshold. Attackers are now operating at machine speed, using AI agents to perform reconnaissance, write exploit code, test vulnerabilities, harvest credentials, and exfiltrate data faster than any human team could match.

The attackers executed roughly 80-90% of their campaign with minimal human intervention. While your security team is in their triage meeting, AI-powered attackers have already mapped your infrastructure, identified critical systems, tested vulnerabilities, harvested credentials, established persistence, and begun exfiltrating data.

This isn't a future threat. This happened in September 2025—less than two months ago. And if it happened once with public disclosure, it's happening elsewhere right now without detection.

If you're still relying on manual incident response processes, you're not just outpaced. You're operating in a completely different time dimension than your adversaries.

The End of Human-Speed Incident Response

For the past two decades, incident response has been fundamentally human-driven. The core response loop has always run at human speed: alert fires, analyst investigates, team convenes, decisions get made, actions get executed. Even in mature organizations, this cycle measures in minutes to hours. For most organizations, it's hours to days.

The Anthropic incident reveals what we're up against now. AI-powered attacks operate in continuous loops, making thousands of decisions per second, adapting in real-time. An attack that might have taken a skilled human team weeks can now be compressed into hours or minutes.

By the time your security team finishes their initial triage meeting, an AI-powered attacker has already mapped your infrastructure, identified critical systems, tested vulnerabilities, harvested credentials, established persistence, and begun exfiltrating data. This isn't hypothetical—this is what actually happened.

The Uncomfortable Truth

Let me share what I hear from incident response teams:

"We have great tools, but we're drowning in alerts. By the time we investigate one, three more have fired."

"Our response time from detection to containment is measured in hours. We're proud of that—it used to be days. But I'm not sure hours is fast enough anymore."

They're right to be worried. The Anthropic disclosure describes attacks moving at thousands of actions per second. Even elite human incident response teams operate at maybe dozens of actions per hour. That's not a small gap—it's orders of magnitude.

When attacks move at machine speed, critical containment windows close instantly. Attack patterns evolve faster than analysts can adapt—the Anthropic attackers broke operations into small, seemingly innocent tasks that looked benign in isolation. Human analysts reviewing alerts sequentially can't pattern-match at that scale.

Fight AI With AI

Here's where the conversation gets uncomfortable. When I tell security leaders that the solution to AI-powered attacks is AI-powered defense, I often get skeptical looks. We've heard "AI will solve security" promises for years, and most haven't delivered. Security teams have AI fatigue before they've even deployed anything useful.

But something fundamental has changed in the past year. The AI capabilities that exist today are qualitatively different from the narrow machine learning models of the past. Modern AI agents can handle complexity, ambiguity, and novel situations in ways that weren't possible before.

The same capabilities that enable AI agents to execute sophisticated attacks can be deployed for defense. Anthropic makes this point directly in their disclosure: "The very abilities that allow Claude to be used in these attacks also make it crucial for cyber defense." In fact, they used Claude extensively to analyze the enormous amounts of data generated during their own investigation of the campaign.

This isn't abstract theory. It's practical reality. Agent-based AI systems deployed today can:

• Monitor at machine speed: Continuously analyze logs in real-time, correlating events across systems and identifying suspicious patterns as they emerge. Not just collecting logs, but actively reasoning about them. Thousands of analyses per second, matching the operational tempo of attacks.
• Execute autonomous response: When threats are detected, AI agents immediately isolate affected systems, terminate malicious processes, revoke compromised credentials, and deploy countermeasures—in seconds, not hours. The response loop collapses from "detect, notify, convene, decide, execute" to simply "detect and execute."
• Adapt to novel patterns: Unlike signature-based detection, AI agents reason about behavior they haven't seen before. When attackers use new techniques or tools, the AI doesn't need a signature update—it recognizes malicious intent from context and behavior. This is critical when facing AI-powered attacks that can adapt and evolve in real-time.
• Operate across unlimited scope: One AI agent can simultaneously monitor and respond across thousands of systems, multiple environments, different cloud providers, and on-premise infrastructure. Scale doesn't degrade response capability. In fact, the AI gets better at identifying patterns as it sees more data.
• Generate real-time intelligence: While responding to incidents, AI agents automatically document what's happening, extract indicators of compromise, map attack progression, and provide detailed forensic analysis—work that traditionally takes human analysts days to complete.

Real-Time IR: A Paradigm Shift

What I'm describing isn't incremental improvement—it's a paradigm shift from reactive to truly real-time response.

Traditional IR is fundamentally reactive. Something bad happens, gets detected, triggers an alert, and then humans react. There's always a gap between compromise and response. Even the best-run security operations centers have this latency. It's inherent to human-driven processes.

Agent-based real-time IR eliminates that gap. Detection and response become a single continuous loop. The system that identifies malicious activity immediately executes countermeasures. Time from "something is wrong" to "threat is contained" drops from hours to seconds. The response isn't triggered by detection—it's integrated with detection.

This changes the calculus for attackers fundamentally. When both offense and defense move at machine speed, attackers lose their temporal advantage. The comfortable operational window they've enjoyed for years suddenly closes.

Consider the Anthropic campaign with AI-powered defense in place: When reconnaissance begins, defensive AI immediately detects anomalous access patterns and rate of queries. Before the attacker finishes mapping infrastructure, containment actions are underway. When vulnerability testing starts, defensive AI recognizes the behavioral pattern and isolates affected systems. The exploit code the attacker's AI writes can't execute because its targets are already quarantined. When credential harvesting attempts begin, defensive AI automatically rotates credentials, revokes suspicious sessions, and flags accounts for verification—all before the attacker can leverage the harvested credentials.

The attack doesn't become impossible. Sophisticated attackers will still find ways in. But it becomes exponentially harder. The attacker's operational tempo advantage disappears. Their sophisticated AI agent burns cycles fighting countermeasures instead of achieving objectives. The cost-benefit equation shifts dramatically in the defender's favor.

What This Looks Like in Practice

Real agent-based incident response doesn't just automate existing playbooks. It provides autonomous reasoning across the entire response lifecycle.

The AI agent continuously monitors all security telemetry—network traffic, system logs, authentication events, application behavior. It doesn't just pattern-match against known signatures; it reasons about behavior, context, and intent. When it sees unusual activity, it doesn't simply flag an alert. It investigates, gathering additional context to determine if it's malicious or benign. It understands your infrastructure, knows which systems are critical, and can assess actual risk.

It automatically triages incidents without waiting for human decisions. Not all alerts are equal, and the AI makes that determination itself, escalating genuine threats and dismissing false positives. When real threats are identified, the agent executes appropriate containment automatically—isolating hosts, blocking IPs, revoking credentials, terminating processes. These actions happen in seconds, not after a meeting.

While containing the immediate threat, it simultaneously investigates scope. It traces the attack path, identifies affected systems, extracts indicators of compromise, and maps the full extent of the incident. It generates the forensic timeline that human analysts would spend days creating. If initial containment doesn't work or the attack adapts, the agent adjusts its approach dynamically. It learns from what it observes, tries different countermeasures, and persists until the threat is neutralized.

For routine threats, it handles everything autonomously. For complex situations requiring human judgment, it presents clear options, explains its reasoning, and executes decisions once made. The agent doesn't replace security teams—it augments them, operating at machine speed so humans can focus on strategic decisions rather than tactical execution.

Building Trust in Autonomous Response

The biggest barrier isn't technical—it's psychological. Security teams are understandably reluctant to let AI make critical decisions autonomously.

But the alternative—human-only response to machine-speed attacks—isn't viable. If attacks execute thousands of actions per second and your response requires human approval at every step, you've already lost.

The solution is graduated autonomy with proper guardrails: Start with AI making recommendations while humans approve. As trust builds, let AI handle low-risk actions autonomously while escalating high-impact decisions. Build in explainable decisions, fail-safe mechanisms, and continuous monitoring of the AI's own behavior. Don't deploy across your entire environment on day one—start with limited scope, prove value, then expand.

Trust is earned through demonstrated value, not demanded upfront.

The Urgency Is Real

Anthropic's disclosure describes a campaign from September 2025—less than two months ago. That's just the one campaign that got detected and publicly disclosed. How many similar attacks are happening right now undetected?

The AI capabilities that enabled this attack aren't proprietary. Multiple AI models have similar agentic capabilities. The techniques are replicable by any sophisticated attacker. We're not facing a future threat—we're facing a current reality most security programs aren't prepared for.

What to Do Now

If you're a security leader, here's my advice:

• Acknowledge the threat. Machine-speed attacks aren't theoretical. They're happening now and accelerating. Your current response capabilities, no matter how good, aren't sufficient.
• Experiment with agent-based IR. Start small—pilot an agent-based system in limited scope. Learn what works and build organizational muscle memory for working with AI defenders.
• Measure your response tempo. Track time-to-detect and time-to-contain for real incidents. Be honest about whether those timelines work against an attacker making thousands of decisions per second.
• Invest in AI for both detection and response. Many organizations have AI-powered threat detection but still rely on human-executed response. That's half the solution. Detection is useless if response takes hours while attacks take seconds.
• Prepare your team. Agent-based IR changes analyst roles from execution to oversight, from tactical response to strategic guidance. Start preparing your team now.

The Inflection Point

Anthropic argues we've reached an inflection point where AI has become genuinely useful for cybersecurity operations—for both attackers and defenders. They're right.

But there's a dangerous lag between when capabilities become available and when they get widely adopted. Attackers are already using AI agents at scale. Most defenders are still evaluating, piloting, or waiting.

The future of incident response isn't human teams working harder. It's AI agents working at machine speed, continuously monitoring, instantly responding, and collaborating with human experts on complex decisions.

That future is already here. The question is whether your organization will adopt it before or after you experience a machine-speed attack.

Given the stakes, I know which option I'd choose.

As a founder building in this space, I'm not neutral about the value of agent-based IR. But the threat described in the Anthropic disclosure is real, and the inadequacy of human-speed response is undeniable. Whether the solution comes from my company or others, the industry needs to move faster. Because our adversaries already have.