AI and Computer Security: The Impending Doompocalypse

Finding vulnerabilities at computer speed doesn't make you secure. It moves the advantage to attackers and increases breach volume.

But you wouldn't know that from watching the market. For the umpteenth time in as many weeks, frontier-lab headlines have pushed security vendor valuations down again. Anthropic and OpenAI are shipping model updates that can spot vulnerabilities in widely used open-source libraries, and public-market investors are reacting as if "AI finds bugs" means the security industry is worth less.

After 25 years in security, I can tell you: finding vulnerabilities has always been the easy part, especially in open source. The hard work is eradication. That means finding every instance across your environment and coordinating the teams to fix it.

The gap between identification and eradication is exactly where breaches happen. And right now, AI is making that gap wider, not narrower.

Don't want to read the whole article? Enjoy this AI generated overview

Attackers Adopt Faster Than Defenders

This shouldn't surprise anyone, but it needs to be said plainly: attackers are adopting AI faster than defenders.

Attackers don't have change-control boards. They don't file tickets. They don't wait for budget approval or coordinate across six teams to deploy a fix. They don't argue over whether security patching is higher ROI than rolling out new features. They find a weakness and exploit it. AI just lets them do that faster.

And these aren't hypotheticals. In just the last few months:

• In November 2025, Anthropic disclosed what it called "the first documented case of a large-scale cyberattack executed without substantial human intervention." A Chinese state-sponsored group designated GTG-1002 jailbroke Claude Code by posing as employees of a cybersecurity firm doing "defensive testing," then used it as an autonomous attack agent targeting roughly 30 organizations worldwide. According to Anthropic's own reporting, Claude executed 80 to 90 percent of all tactical operations independently. (Anthropic, The Hacker News, Axios)
• In March 2026, a CMS misconfiguration at Anthropic left roughly 3,000 unpublished internal assets publicly accessible, exposing the existence of an unreleased frontier model called Mythos. The leaked materials described the model as a "step change" in capability and warned of "unprecedented cybersecurity risks." Anthropic attributed the exposure to human error. (Fortune, Techzine)
• Five days later, Anthropic shipped an npm package containing a 59.8 MB source map file that exposed approximately 512,000 lines of Claude Code's TypeScript source across 1,906 files. The root cause? A missing *.map entry in .npmignore. Within hours the code was mirrored on GitHub, and Anthropic's subsequent DMCA takedowns accidentally removed roughly 8,100 legitimate repositories before being narrowed. Zscaler's ThreatLabz team later found threat actors seeding trojanized Claude Code versions distributing Vidar Stealer and GhostSocks malware. (The Register, BleepingComputer, Zscaler, TechCrunch)

Think about what that sequence tells you. The organizations building these models, the very people who are supposed to understand this technology better than anyone on earth, can't keep their own house secure. A missing line in .npmignore exposed half a million lines of source code. A CMS misconfiguration blew the cover on an unreleased model. A nation-state group turned their own product into an autonomous weapon.

If frontier LLMs were security solutions rather than security tools, the people who build them would be the most secure organizations on the planet. They're not. Nobody is.

LLMs are tools, not solutions. And right now, the bad guys are using them better than most defenders.

From Identification to Eradication: What Most People Don't Understand

It's painfully clear that most people outside of security don't understand what it takes to "do security" inside a real company. Every single vulnerability follows a workflow that looks something like this:

1. Awareness and Vulnerability Intelligence

The first challenge is simply knowing a vulnerability exists. In 2025, 48,185 vulnerabilities were disclosed via the Common Vulnerabilities and Exposures (CVE) program, about a 20% increase over 2024. Roughly 40% of the reported vulnerabilities were assigned a critical or high rating.

In 2026 so far, approximately 28,450 CVEs have been reported. That works out to about 287 new vulnerabilities. Every. Single. Day.

Yes, LLMs can help identify vulnerabilities faster. But faster identification without faster eradication just means a longer list of things you can't fix quickly enough.

2. Inventory and Scoping

Next, you need to understand if and where the vulnerability might be present in your environment. This is harder than it sounds:

• Cloud-native sprawl. Containers spin up and down. Ephemeral compute disappears before you can scan it.
• Faster release velocity. LLMs write code at breakneck speed and CI/CD pipelines push changes daily or hourly.
• Transitive dependencies. A package depends on a package depends on a package, until one of them four layers deep turns out to be vulnerable.

You can't patch what you can't find. And most organizations can't find everything.

3. Ownership and Attribution

Here's what you don't know about security professionals: we spend most of our time figuring out who needs to fix the problem, not finding the problem itself.

After we identify who owns the system or code, we barter with leadership for the human time to implement the fix. We're competing against roadmaps, feature releases, and quarterly goals for developer hours that are already overcommitted.

This is where the real work happens. Not in detection. In negotiation.

4. Attack Path Analysis

Not all vulnerabilities are exploitable in your environment. Before burning political capital to get something fixed, you have to assess exploitability based on your specific implementation context, adjust risk (because not all CVEs are equal and context changes everything), and produce evidence that developers and admins will actually accept.

They want proof before they trade remediation time for feature work. And honestly? I don't blame them. If I'm a developer and a security engineer tells me to drop everything for a vulnerability that isn't actually reachable in our environment, that's a credibility problem.

5. Patching and Remediation... or Not

After you confirm impact, find the owner, and win the prioritization fight, you finally get to patching. But it's not always that simple. Some vulnerabilities don't have patches. You need compensating controls, mitigations, validation, and monitoring. This can take weeks. Sometimes months.

Meanwhile, attackers with AI-powered scanners have already moved on to the next target: you.

LLMs Aren't Security Solutions: They're Components

LLMs don't ship code by themselves. Harnesses like Claude Code and Codex do. Same story in security: LLMs are great at discrete tasks, but only when wrapped in purpose-built workflows and tooling. That's what security vendors are building.

There's a reason coding was the first major LLM use case to mature. There are massive volumes of training examples, fast and objective feedback loops (compile, run, pass/fail), and millions of users generating reinforcement signals every day.

Security doesn't have those properties. You can't run eradication end-to-end without humans. The problem space isn't just vulnerabilities. It's misconfigurations, business logic flaws, human error, and attacker creativity. The threat landscape changes constantly. The technology landscape changes constantly.

Defenders have to be right every day. Attackers only have to be right once.

That's exactly the operating environment where LLMs struggle to work independently.

Why Frontier Labs Won't Build Security Harnesses

Maybe frontier labs will adapt coding harnesses to produce more secure code. But code security is one discipline in a much larger security ecosystem.

Building an EDR platform isn't a model capability problem. It's a decade of kernel-level engineering, customer deployment infrastructure, regulatory compliance across dozens of jurisdictions, and 24/7 incident response relationships that took years to build. A SIEM isn't a language problem: it's a data pipeline engineering challenge refined over years of operating at the scale of billions of events per day. Compliance monitoring requires deep, constantly-updated regulatory domain knowledge across industries and geographies.

Frontier labs aren't going to staff SOCs. They aren't going to build the go-to-market muscle to sell, deploy, and support security tooling inside ten thousand enterprises. They're going to partner with (or sell capabilities to) the companies that already have that infrastructure.

The Economics of Security Data

Here's where the "AI replaces security vendors" thesis falls apart completely.

100 employee workstations can generate roughly 300,000 events per hour. That's just endpoints, before you add servers, cloud workloads, network devices, identity systems, and SaaS applications.

Let's do the math. A typical security log event in JSON format runs about 300 tokens. At 300,000 events per hour, that's 90 million input tokens per hour for just 100 machines. Even using the cheapest current-generation frontier model, Anthropic's Haiku 4.5 at $1 per million input tokens, you're looking at $90 per hour just for ingestion. With a mid-tier model like Claude Sonnet 4.6 at $3 per million input tokens, it climbs to $270 per hour. And that's before output tokens, which cost 5x more ($15 per million for Sonnet, $5 per million for Haiku). Factor in even modest output for analysis summaries and you land around $6,500 per day for 100 workstations at Sonnet pricing.

Now scale that to something real. A mid-size enterprise runs 5,000 endpoints, servers, and cloud workloads. That's 50 times the load: over $300,000 per day. And remember, endpoints typically represent about 20% of the total attack surface. Want to cover network telemetry, identity logs, cloud API calls, and SaaS activity too? You're staring at seven figures. Per day. For data that yields actionable findings maybe 0.1% of the time.

LLMs don't scale to raw security data. They never will at these economics, and they don't need to.

What actually works, what security teams and vendors have been refining for decades, is pre-processing:

• Filter common and benign events
• Minimize logs to relevant artifacts
• Group events by threat relevance
• Send the reduced set to analysis, now potentially LLM-powered

LLMs get incredibly useful after the data is shaped. They can interpret complex event chains, correlate across data sources, and surface insights that would take a human analyst hours. But they need the data engineering layer underneath them. The vendors who have spent decades solving this exact problem are the ones positioned to harness LLMs effectively. The AI lets them extract insights faster once the data is prepared correctly. That's a multiplier on existing capability, not a replacement for it.

Security Vendors Are More Valuable Than Ever

Let's be direct about what's coming.

Security vendors will respond to and remediate the onslaught of LLM-identified vulnerabilities. They'll block attackers using LLMs to find and exploit weaknesses faster. They'll provide the workflows that let humans harness LLMs safely. And they'll be on the front lines with customers when things go sideways, because things will go sideways.

The best vendors will be the ones who cost-effectively harness LLMs to 10x their practitioners, so small teams can stay ahead of attackers who are already using the same technology.

The Real Risk Calculus

Here's what the market is getting wrong:

• AI makes vulnerability identification faster, for everyone, including attackers
• The identification-to-remediation gap is where breaches happen
• That gap widens when attackers have AI and defenders don't have the tooling to keep up

AI is a force multiplier, not a replacement. The hard part has never been finding problems. It's fixing them at scale, under pressure, with limited resources.

Are you ready to bet against the people who've been doing exactly that for decades?