Before Anthropic’s recent code security announcement tanked the market, I was lucky enough to be at a private event with the company’s Deputy CISO Jayson Clinton and the ex-OpenAI CISO Matt Knight, who built OpenAI’s vulnerability detection product Aardvark) Our discussion wasn’t just about what’s in the new AppSec stack, but whether there will there be more or fewer security engineers in the future.
I run Semgrep, a code security company that has been working at the intersection of developers and security for years. I think there’s no future for AppSec without LLMs, and I have a lot of genuine excitement for this moment. We have a generational opportunity to make software construction more secure. But there are some important questions that remain unanswered.
Anthropic isn’t the first company to experiment with finding vulnerabilities with LLMs. Google’s BigSleep agent found and disclosed a real-world memory safety vulnerability in November 2024. OpenAI launched Aardvark, an agentic security researcher, in October 2025. The race to apply AI to vulnerability detection has been underway for more than a year.
Will AI Replace Security Engineers?
No, but the role is changing. Emerging AI agents act as virtual security engineers by handling the high-volume, repeatable work of running scans, triaging findings, and suggesting patches. Human security professionals are moving into AppSec manager roles, where they focus on strategic security work, review edge cases and make judgment calls on ambiguous findings.
AI in Cybersecurity Is Missing Some Numbers
But so far, none of the foundation model companies (BigSleep, Aardvark, OpenAI) have published detailed statistics on how many false positives they experienced to get the results they had or the cost to do so. Was this a $1M investment? $10M? What were the false positives rates, and how many human hours of review time were required? All that information matters.
Why does it matter? DARPA’s AI Cyber Challenge competition, one of the most rigorous public benchmarks for AI in security, required all participants to disclose the cost per vulnerability as well as the confusion matrix of false positives. That standard exists because, without it, you can’t compare tools, you can’t make procurement decisions and you can’t understand whether what you're seeing is a breakthrough or a well-resourced demo. Perhaps it is now CPUs and GPUs (with fuzzing and LLMs) rather than human eyeballs that “make all bugs shallow,” but how much will that cost relative to an organization’s security budget?
What Actually Makes LLMs Effective at Security?
LLMs are genuinely powerful tools for security work. But in our team’s experience, giving LLM security agents access to deterministic analysis tools makes them dramatically more effective, particularly when it comes to false positives. A raw LLM reasoning over source code is working largely from pattern recognition and inference. That produces a lot of noise.
But an LLM agent that can invoke a precise static analysis tool to verify its hypothesis (to check whether a data flow actually reaches a sink or whether an authorization check genuinely fires, etc.) produces a much higher-quality signal. It’s also cheaper, which is why the cost-per-vulnerability metric matters so much. A tool-augmented LLM agent should, in principle, spend less compute chasing false leads, which means lower cost and higher precision at scale. No one wants to spend $10K per scan of their monorepo!
If LLMs are the new virtual security engineers, they will benefit from great tools the same way human security engineers do.
Virtual Security Engineers, Human AppSec Managers
What’s emerging is a model where AI agents act as virtual security engineers: They run scans, triage findings, suggest patches and handle the high-volume, repeatable work that currently consumes most of a security engineer’s time. Humans, in turn, move into an AppSec manager role: setting priorities, reviewing edge cases, making judgment calls on ambiguous findings and focusing on the strategic security work that requires organizational context and human judgment.
This is already happening. Security teams that have deployed AI-powered triage are reporting dramatic reductions in manual workload and, for the first time, making real progress against vulnerability backlogs that had been growing for years. The key is human oversight remaining in the loop and the ability to mechanically explain decisions made by LLMs.
A Call to the Model Companies
What I’d ask of every AI lab shipping security tools: publish your confusion matrices and cost curves. The labs that do this will build more trust, accelerate real-world adoption, and help set a standard that raises the floor for the entire industry.
