When Hackers Can Attack with AI in 22 Seconds, Why Is Enterprise Defense Always a Step Behind?

Event Background

Google's Mandiant cybersecurity intelligence unit disclosed in its *M-Trends 2026* report that hackers have reduced the time from "initial breach to handoff to the next wave of attackers" from 8 hours in 2023 to just 22 seconds in 2026. This timeframe falls far short of typical enterprise security analyst response chains: alert detection → threat identification → management notification → incident response activation—a process that usually requires 3-5 minutes.

In the same period, enterprises enthusiastically embrace generative AI and autonomous agents, entrusting sensitive data, business processes, and transaction permissions to these "automated decision-makers." Attackers are doing exactly the same: using large language models to generate high-fidelity phishing emails, automating reconnaissance, and embedding AI into malware to adapt to defensive measures.

The result is: *enterprises use AI to accelerate business processes, attackers use AI to accelerate attack processes, yet traditional security operations led by human decision-making have not accelerated at all*.

Key Observation: The Time-Gap Paradox

This is not simply a problem of "defense technology being insufficiently advanced." Even if enterprises deployed identical AI defense tools tomorrow, they would still face a structural dilemma:

1. Mismatch Between Decision Time and Execution Time - Attacker AI: "Detect vulnerability → Generate phishing email → Auto-execute," end-to-end 22 seconds - Enterprise defense process: "Detect threat → Confirm risk → Seek management approval → Execute isolation," end-to-end 3-5 minutes - The order-of-magnitude difference stems not from technological gaps, but from decision authorization structures

2. Symmetric Technology, Asymmetric Constraints - Decisions attackers can make in 22 seconds (infiltrate, move, escape) require enterprises to navigate legal, compliance, and risk control frameworks that may necessitate longer approval chains - This is not something AI can directly accelerate—approval cannot compress from 5 minutes to 22 seconds unless the enterprise accepts substantially greater risk

3. The Self-Imposed Paradox of "Agentic Defense" - Google proposes "Agentic Defense" as a solution: use AI agents to automate first-line response - But this introduces new risks: enterprises are already using AI agents to manage sensitive data and business processes; now they must use AI agents to defend against attacks on those very agents - It's like "using poison to cure poison"—the likelihood that the defense infrastructure itself becomes compromised by attackers is also rising

Core Principle: The Structural Advantage of the OODA Loop

Military strategist John Boyd introduced the "OODA Loop" (Observe → Orient → Decide → Act), arguing that victory in combat depends not solely on resources or firepower, but on whose decision cycle moves faster. When one side's loop is faster than the other's, the opponent is perpetually reacting to the previous action, locked in a passive stance.

Cybersecurity is evolving into an OODA Loop competition: - Attacker's loop: Probe → Generate attack payload → Execute → Observe defense response → Adapt quickly - Because attackers need no prior approval, the loop is extremely tight - Each iteration requires only 22 seconds

• Defender's loop: Detect → Analyze risk → Seek decision-maker approval → Execute isolation → Assess effectiveness
• - The loop embeds multiple layers of human approval
• - Each iteration requires 3-5 minutes, often longer

When the attacker's loop is 8-15 times faster than the defender's loop, the defense side can never catch up.

Why This Cannot Be Solved With "Faster AI" Alone

Some would argue: "Then enterprises should deploy AI defense agents too; they could also respond in 22 seconds." This logic overlooks three things:

1. Trust Deficit - Enterprises dare not let AI completely autonomously decide "which systems to isolate" or "which users to block" - Because the cost of misjudgment (business interruption, regulatory fines) is fatal to enterprises, but to attackers it's merely "try again" - So even with AI assistance, the human approval layer is difficult to eliminate entirely

2. Asymmetric Objectives - Attackers need to succeed just once to achieve their goal - Defenders must succeed 100% of the time - This means defense must retain more redundancy and caution, unable to be as aggressive as attack

3. Asymmetric Complexity - Attacks can be optimized for a specific target (precision strike) - Defense must protect against all possible attack surfaces simultaneously (comprehensive defense) - Thus the defender's decision tree is always more complex than the attacker's

Historical Analogy

This dilemma is not new. It echoes many patterns seen in military and business competition:

• Nuclear Arms Race During the Cold War: Both the US and USSR accelerated missile response times until reaching "decide victory within 15 minutes of launch," making any diplomatic negotiation impossible. The acceleration of technology itself increased conflict risk.

• High-Frequency Trading vs. Regulation: In financial markets, high-frequency trading algorithms execute thousands of transactions in milliseconds, while regulators require hours or days to respond. The result is frequent "black swan events" (such as the 2010 Flash Crash), yet root causes cannot be prevented in advance.

• Social Media Misinformation vs. Content Moderation: The spread speed of false information (reaching millions in minutes) vastly exceeds platform moderators' response capacity (hours to days), meaning moderation is always "closing the barn door after the horse has bolted."

These cases all point to the same conclusion: *when one side's "execution speed" vastly exceeds the other side's "decision speed," accelerating technology itself cannot change the inevitable outcome of failure*.

Implications for Enterprises

If enterprises cannot compress their decision cycle from 5 minutes to 22 seconds (which is virtually impossible within legal and risk control frameworks), then in this "defense-attack acceleration" arms race, enterprises have already lost—even if they spend a billion dollars today deploying the most advanced AI defense tools.

The only way out is not to accelerate technology, but to change the playing field of the confrontation:

• Reduce the attack surface: Do not let AI agents autonomously manage sensitive data. Data permissions should remain in human hands.
• Raise the barrier to entry: Through architectural design, make initial compromise itself difficult (such as zero-trust architecture, multi-factor authentication), rather than relying on post-incident rapid response.
• Accept the time lag in defense: Acknowledge that you can never respond in 22 seconds, so design "buffer layers"—even if attackers succeed in entering, they remain trapped in isolation zones unable to cause real damage.

In other words, *the answer to asymmetric acceleration is not "faster," but "thicker"*.