AI can turn a published security patch into a working attack within hours. Why monthly updates are now too slow, and what small businesses can do.
- Date
- Reading
- 5 min read
- Written by
- Nick Li
- Tags
- #AI#Risk
For the last few years, AI has mostly shown up in cyber security as a solution: automatic scans, automatic alerts, automatic log analysis. This year's research shows the other side. The same ability can help an attacker turn a vendor's fresh security patch into a working attack within hours. Patching is becoming a race, and most small businesses are still running it at a once-a-month pace.
Every patch is a map for attackers
When a software company fixes a security hole, it releases a new version. Compare the old code with the new code and you can see what changed, and from that work out where the original flaw was. This is called patch diffing, and it used to take a skilled specialist days or even weeks.
In June, Anthropic published a study testing how fast AI can do this:
- Firefox: from 18 Firefox patches, the unreleased Mythos Preview model built 8 attacks that could run code. The first took about an hour, while the fixed Firefox release was still 18 days away.
- Windows: from 21 Windows kernel flaws, it built crash proofs for 18. The first took 31 minutes, and all 18 were done within six hours. Each full attack that took over the machine cost about US$2,000 to produce.
- Public models managed some too: models anyone can use today also built working attacks, just fewer of them.
The researchers' conclusion is blunt: one person can now turn a month's worth of patches into working attacks in a single afternoon, for a few thousand dollars and with no specialist knowledge.
What the numbers say
This isn't confined to the lab. Mandiant's M-Trends 2026 report, published in March, puts the average time to exploit a vulnerability at about minus seven days: many flaws are being used before a patch even exists. Exploiting vulnerabilities was the most common way in for the sixth year running, at 32% of intrusions.
The number of flaws is climbing fast, too. Rapid7's quarterly report counted 8,539 high and critical vulnerabilities in the second quarter of 2026, double the 4,268 a year earlier. As SecurityWeek quotes the report, traditional patch cycles are being overwhelmed by the sheer volume of vulnerabilities and attacker speed.
AI helps defenders too, but people still do the fixing
To be fair, AI is helping the defending side as well. In April Anthropic launched Project Glasswing, giving partner organisations and open-source projects the same model to look for flaws. It has found thousands of high-severity vulnerabilities.
And that is exactly the problem: finding flaws has sped up, fixing them hasn't. In its April technical report, Anthropic said fewer than 1% of the flaws found so far had been fully patched. A fix needs someone to understand the problem, change the code, test it and schedule the update, and every step needs people. Finding bugs can be handed to AI, but getting every server and website updated still takes time and hands.
Why small businesses lose out most
Large companies have security teams, test environments and overnight update windows. In a small business it often looks like this:
- The website was built years ago by an outside agency, and nobody has updated it since.
- WordPress plugins, the NAS, the router and the accounting software are each updated by a different person, when they have time.
- Updates wait for the IT person's end-of-month visit, or until something breaks.
There used to be a buffer of weeks between a patch coming out and attacks appearing. Now it may be hours. Updating once a month means leaving a publicly known hole open on the internet for up to a month.
If you can't match the speed, change the game
A small business can't out-staff attackers. It can shrink what it has to keep up with:
- Turn on automatic updates wherever you can. Browsers, computer and phone systems, and WordPress security releases can all update themselves. Anthropic's study points to Firefox's automatic background updates as the model to follow.
- List what you run. Write down every system the company uses, its version, and who updates it. If you don't know your versions, you can't tell which security notice applies to you. We show how in October 2026 security updates: which versions to install
- Put less on the internet. Remove unused plugins and old websites, and don't leave admin pages open to the whole internet. A static website has no database or plugins to keep patching, one of the reasons we recommend static sites.
- Set deadlines for updates. Our suggestion: for systems that face the internet (website, VPN, email server), handle a critical flaw within 48 hours; check everything else weekly.
- Ask your suppliers. If an agency maintains your website or system, ask who handles security updates and how fast. Start with the questions in A PHP security flaw: what HK SMEs must ask today.
- Prepare for the worst. Test your backups regularly and keep at least one copy offline, so that if you do fall behind you can recover quickly.
Patching used to be a job you could schedule. Now it's a race.
AI itself isn't the thing to fear. It has made both attack and defence much faster. The real risk is meeting a new opponent at the old speed. If you'd like to check where your company's systems and website might be falling behind, let's talk.
