AI Is Getting Better at Finding Security Holes. Here's What That Means.

TL;DR: The big AI labs are now using their own systems to find security weaknesses in real, widely used software — and to help get those weaknesses fixed. That is good news if the fixes reach people before anyone can misuse them, and a reminder for the rest of us to keep our software up to date.

For a long time, finding a serious hole in a piece of software was slow, specialised work. That picture is changing, and quickly. Over the past year, Anthropic and OpenAI have shown that their AI systems can find genuine weaknesses in real software, including programs already checked by experts for years.

This is a plain-English overview of what they have shared, why it matters if you run a business or just use a computer, and a few simple takeaways. It is not a technical guide, and it is not a reason to panic.

What Anthropic has been doing

In September 2025, Anthropic’s security researchers wrote that AI had reached a turning point: the systems were becoming genuinely useful for this kind of work, and the people who defend software needed to start using them in earnest. You can read that post here: Building AI for cyber defenders.

By February 2026 they had put that idea into practice. Using Claude Opus 4.6, they found and confirmed more than 500 serious weaknesses in widely used open-source software — the kind that quietly sits underneath a huge amount of the internet, often looked after by small teams or volunteers. Some of those issues had gone unnoticed for decades. They began reporting them to the people who maintain the software, and the first fixes started to land. That’s in their February 2026 research post.

A couple of weeks later, Anthropic released Claude Code Security as a limited preview. Teams can point the same kind of review at their own software, get a list of possible problems with suggested fixes, and have a person decide what to change. Nothing is applied automatically.

The work has continued through the year. In late October 2026, Anthropic described further research in which an internal system found serious issues in some of the world’s most widely used software, including web browsers and operating systems, and reported them so they could be fixed. See their late October research post.

What OpenAI has been doing

OpenAI has been on a similar path, and has been quite open about naming the tools.

In October 2025 it introduced Aardvark, an AI helper designed to look through software the way a human security researcher might: read the code, try to confirm a problem is real, and suggest a fix for a person to review. In early testing it also found issues in open-source projects, which OpenAI reported through the usual responsible channels. In March 2026, Aardvark was folded into OpenAI’s coding tools and renamed Codex Security.

Then, on 5 November 2026, OpenAI introduced GPT-6 Cyber: a model trained specifically for this kind of security work. According to OpenAI, it is already being used to find and help fix problems in real software. One example they shared is work with Mozilla: the model helped uncover serious issues in Firefox, and Mozilla has already shipped fixes for many of them.

Why this matters

There is a simple race underneath all of this. The same kinds of AI systems that help a software team find a weakness can, in the wrong hands, help someone look for a way in. Anthropic and OpenAI both say this out loud. Their argument is that there is a window now: use these tools to find and fix problems faster than they can be misused, and get those fixes out to the people who run the software.

If you run a small business, you may never touch the tools themselves. You still live with the results. The programs you and your customers use every day — browsers, operating systems, the hidden pieces that websites and payment systems rely on — are part of what this research is aimed at. When a fix is released, the benefit only reaches you if you actually install it.

There is a human piece too. In the write-ups from both labs, a person still checks the findings and approves the fixes. AI can scan a lot of software quickly. Judgement about what to change still belongs with people.

A few simple takeaways

Keep your software up to date. This is the unglamorous one, and it matters more than it used to. Many of the issues these labs are finding turn into ordinary updates for browsers, phones, computers and the tools you already use. Turning on automatic updates, and not putting off the ones that ask you to restart, is the most useful thing most of us can do.

Treat “later” as a risk, not a reminder. An update sitting uninstalled is a known problem with a fix already sitting in the box. Install it when you can.

If you pay someone to build or look after your software, ask a plain question: how do you find security problems, and how quickly do fixes get applied? You do not need a technical answer. You want to know that someone is looking, and that updates are not left on a shelf.

Don’t assume a small business is too small to matter. A lot of the software underneath small firms is the same software everybody else uses. When a weakness is found in a common tool, it can affect a lot of people at once. Staying current is how you stay in the group that is already patched.

The short version is this. AI is now good enough to help find security holes in the software the world already depends on, and the major labs are trying to point that ability at defence. The rest of us do not need to become security researchers. We do need to let the fixes through the door.