AI vs. AI: How Forty Years of Cybersecurity AI Led Us to the Real Bottleneck
TL;DR: For forty years, AI in cybersecurity was about finding the needle in a haystack. Today, AI finds thousands of vulnerabilities every week. With attackers operating at machine speed, the real security bottleneck has shifted from detection to remediation. To stay protected, organizations must measure time-to-patch vs. alert volume, deploy defensive AI, and treat every AI agent as a privileged identity.
—
For forty years, the fundamental promise of AI in security was simple: help us find the needle in the haystack.
Today, that problem is essentially history. The new reality is that AI now finds thousands of needles every single week, and almost no organization can pull them out fast enough.
To understand how we arrived at this moment, it helps to look at how we arrived, how the bottleneck shifted, and what mid-size businesses actually need to do about it today.
The Five Eras of Security AI
1. Rules and Statistics (1980s–1990s)
In 1987, Dorothy Denning articulated a foundational idea: model normal behavior, then flag deviations from it. Early intrusion-detection systems relied heavily on expert rules hand-written by analysts. While clever, these systems were inherently brittle and notoriously flooded teams with false alarms. Surprisingly, a large portion of the market still sells variations of this exact concept.
2. Machine Learning Arrives (Late 1990s–2000s)
Machine learning earned its keep in cybersecurity inside our email inboxes. Bayesian spam filters in the early 2000s marked the first mass-market win for security Machine Learning.
However, translating that success to catching network intruders hit a wall. In 2010, Robin Sommer and Vern Paxson pointed out the core dilemma: base rates. Real attacks are rare, meaning even highly accurate models bury analysts in false positives. More importantly, Machine Learning excels at spotting things similar to what it has seen before, whereas security is primarily about encountering what you haven't seen. When a black-box model fires without explaining why, an analyst cannot take decisive action.
3. Behavioral Analytics & Machine vs. Machine (2010s)
The real turning point arrived in 2016 at DARPA’s Cyber Grand Challenge, where machines found, exploited, and patched software vulnerabilities against each other with no humans in the loop. "AI vs. AI" ceased to be a marketing slide and became an operational reality.
Concurrently, defensive models themselves became prime targets:
2017: NotPetya demonstrated the catastrophic potential of supply chain compromises, causing an estimated $10 billion in damage worldwide without using any AI at all.
2019: Researchers at Skylight Cyber bypassed a leading Machine Learning antivirus simply by appending harmless strings to malware files.
2019: Fraudsters used cloned CEO voice audio to trick an executive into transferring €220,000. Every new defensive model quickly grew into a new attack surface.
4. Generative AI (2022–2024)
When ChatGPT debuted in late 2022, attackers upgraded far faster than defenders. Criminal chatbots emerged within months, erasing typos and bad grammar that used to be reliable indicators of phishing. By 2024, deepfake technology reached a tipping point when an engineering firm lost roughly $25 million after an employee joined a video call where every other attendee was synthetic.
Defenders largely received "Copilots," which are tools adept at summarizing alerts, but far less capable of deciding how to resolve them. Crucially, our own AI systems became the primary target, with prompt injection emerging as SQL injection’s chattier younger sibling.
5. Autonomous Agents (2025–Today)
Over the past eighteen months, AI transitioned from advising humans to acting independently:
August 2025: Finalists in DARPA’s AI Cyber Challenge independently identified 86% of planted vulnerabilities and patched most of them without human intervention.
November 2025: Anthropic disclosed an espionage campaign where AI carried out 80% to 90% of the operational work.
Spring 2026: Partners in Anthropic’s Project Glasswing utilized the Mythos model to flag over 10,000 high- and critical-severity vulnerabilities in essential software in about a month.
July 2026: Reports highlighted agentic ransomware (such as JADEPUFFER) executing attack chains end-to-end. Whether an agent executes 90% or 100% of an attack autonomously is beside the point—a single human operator now wields the output of an entire crew that never takes weekends off.
The New Bottleneck: Finding vs. Fixing
After forty years, the operational bottleneck in cybersecurity has moved. First it was detection; then it was triage. Today, finding vulnerabilities is close to free, but fixing them still relies on change tickets, approval chains, and maintenance windows.
Early reports indicate that only a small fraction of the thousands of vulnerabilities discovered by AI models like Glasswing have actually been patched. It is the classic Sommer and Paxson base-rate problem returning in a brand-new outfit: uncovering flaws at machine speed creates zero value if remediation moves at human speed.
Three Steps for Mid-Size Companies This Quarter
If your organization doesn't have an army of dedicated security researchers, focus your resources on these three practical priorities:
Emphasize Time-to-Patch, in addition to Alert Volume: If your mean time to remediate critical vulnerabilities is measured in months, that window is your actual security exposure.
Deploy AI in Detection and Response: Attackers are operating at machine speed; defensive workflows must match that cadence to keep pace.
Treat AI Agents as Non-Human Identities: Every AI agent deployed in your environment must be managed as a privileged identity with strict permission boundaries—because that is precisely how an attacker will exploit it.
Your attackers' AI isn't standing idle. The fight isn't won by who identifies the most alerts, but by who closes the loop from discovery to fix the fastest.
Download your free SMB Readiness Checklist here: https://docs.google.com/document/d/1CR1tRgTA1IVD5HqW-WQeCLknetu2Z-8OJE2_oInyPds/edit?tab=t.0
