5 min readRSSFlux

Massive Numbers of Zero-days Are Being Created Because of AI — What Do We Do?

AI has made finding software vulnerabilities cheap and fast, while the price of a working exploit keeps falling. The result is a threat landscape where small operators are targeted directly, and patching on a schedule no longer works.

cybersecurityvulnerabilitiesAIzero-day exploitsself-hostingpatching

Rob Braxman's video opens with a number that is hard to absorb: roughly 129,000 vulnerabilities identified in a single year, in the work he attributes to Anthropic's security research. Whatever the exact figure, the direction is not in dispute. Finding software flaws has become a machine-scale activity, and the security industry's model of "discover, disclose, patch, repeat" is operating at a tempo most organisations cannot match.

The interesting question is not whether AI writes better exploits. It is what happens to everyone else when the cost of finding a flaw collapses.

The numbers are the story

Security work has always been expensive because it was labour-intensive: read the code, reason about edge cases, prove the bug is reachable, write the exploit. Large language models are unusually good at the middle steps. They read code quickly, spot patterns that look like mistakes, and draft proof-of-concept code without getting bored.

Braxman points to a concrete example: a Linux vulnerability scoring 9.8 on the CVSS scale — full remote control of the machine — that was still unpatched in mainstream Ubuntu updates at the time of the video. That is the uncomfortable part. The vulnerability was known, the severity was maximal, and the patch had not reached the systems that needed it.

That gap between disclosure and deployment is where most real-world compromise happens. It is also the part a small operator can control.

Zero-days got cheaper, and that changed who can play

A zero-day used to be a strategic asset. Nation-states paid large sums for them and hoarded them, because a reliable unpatched exploit is only valuable while it is secret.

Two things are now true at once. Intelligence services and criminal groups are still stockpiling vulnerabilities, and the supply has grown enough that the price has fallen. When a working exploit costs less, the set of attackers who can afford one widens from governments and organised crime to mid-level operators who read forums.

That is a different threat model from the one most small businesses plan for. You do not need to be a target worth a nation-state's budget to be worth a $500 exploit.

From mass breaches to individual targeting

The second shift is in targeting. Large database breaches are noisy and increasingly well-defended. The cheaper play is individual: a specific person's mail account, a specific small company's server, a self-hosted service with a known flaw and no patching routine.

For anyone running their own infrastructure, that is the relevant change. Self-hosting removes your data from a platform's visibility, but it also removes the platform's security team. You become the person responsible for the patch window, the firewall rules, and the backups.

Why small operators get hit hardest

The failure mode is rarely exotic. It looks like this:

  • A service is exposed on a public port because that was the easiest way to reach it.
  • A dependency has a known critical vulnerability, and updates are applied monthly, or when something breaks.
  • Nobody is watching advisories, so a 9.8 severity bug is discovered by an attacker before it is discovered by the admin.
  • Backups exist but have never been restored, so recovery is theoretical.

None of these require a sophisticated adversary. They require an adversary who checks whether you patched.

A practical hardening checklist

There is no way to defend against an unknown flaw with certainty. There is a lot you can do to make the known ones irrelevant.

  1. Patch on a schedule, not on a reminder. Automatic security updates for the base OS, and a weekly review for anything that needs a restart. The 9.8 example matters only if you are still running the unpatched version.
  2. Shrink what is exposed. Every public port is an invitation. Move administration and internal services behind a private network such as Tailscale, and close the port entirely.
  3. Watch advisories like you watch your inbox. Vendor security notices, distribution security trackers, and the CISA known-exploited list are all published as feeds. Subscribe to them instead of relying on a news article to tell you.
  4. Reduce your public footprint. Hosting personal services yourself, on hardware you control, means fewer third parties holding your data — and fewer organisations that can leak it.
  5. Use a VPN on untrusted networks. It does not make you anonymous, but it removes a large class of interception and injection attacks on café and hotel Wi-Fi.
  6. Be honest about antivirus. Signature-based scanning catches known malware. It does not catch a freshly written exploit aimed at your server.
  7. Test a restore. An untested backup is a hope, not a plan.

AI cuts both ways

The same tooling that finds vulnerabilities can help defend against them: scanning configuration, summarising logs, flagging anomalous behaviour, and triaging alerts faster than a person can read them. Braxman's framing is that AI is now part of running a server, not an optional extra.

The complication is that AI also writes more code, and more code means more bugs. Speed of generation and speed of review are now mismatched across the entire software supply chain, which is why the volume of vulnerabilities is unlikely to fall.

What to take away

The practical lesson is not "AI is dangerous." It is that the interval between a vulnerability becoming public and being exploited has compressed, so the interval between a patch being published and being deployed has to compress too.

That is an operational discipline, not a product you can buy. But it helps to have the inputs automated — and advisories are exactly the kind of source that benefits from being a feed. RSSFlux turns pages without RSS, including vendor security notices and distribution trackers, into feeds your reader or chat bot can watch. See how it works, or read the X/Twitter feeds note if you follow researchers there.

Source: Massive Numbers of Zero-days (Exploits) are Being Created Because of AI - What Do We Do? — Rob Braxman Tech.

Frequently asked questions

What is a zero-day exploit?
A zero-day is a vulnerability that is exploited before the vendor knows about it, so there is no patch available at the moment of attack. Once a fix ships, exploitation of the same flaw becomes an n-day attack against anyone who has not yet updated.
Why is AI making this worse?
Finding bugs is largely pattern recognition over source code, and that is exactly what large models are good at. AI lowers the cost of discovering vulnerabilities and writing exploits, so more people can do it, and they can do it faster after a patch is published.
What should a small operator actually do?
Shorten the time between patch release and deployment, remove services from the public internet, use private networking for remote access, and monitor vendor advisories as a feed so you hear about issues the same day.

Source

Original article: Rob Braxman Tech

Monitored via RSSFlux — items from this feed are collected, formatted, and shared here automatically. Turn your own feeds into content.

Do this without writing code

RSSFlux turns any public web page into a feed, keeps it fresh, and routes updates to Slack, Discord, Telegram, email, or your own API. Start free.

Related reading

6 min read

Proxmox: The OS You Put on Your Homelab Servers

A practical walkthrough of installing Proxmox VE on a home server and running your first virtual machines — Ubuntu Server, Nextcloud, and a self-hosted AI agent — without exposing anything to the public internet.

home labProxmoxvirtual machinesRead →
5 min read

Is SaaS Dead? Box CEO Aaron Levie on What AI Changes

Box's CEO argues the next software layer is not another app but a context layer: governed access to enterprise data that thousands of agents can query. Here is what that means for tokens, pricing, and open models.

AI agentsenterprise softwareBoxRead →
5 min read

Android is the Better Linux. But Wait ...

Android runs the Linux kernel and ships with hardening desktop Linux has never enforced: mandatory app policy, verified boot, and locked-down runtimes. The catch is who holds the keys.

Android securitydesktop LinuxSELinuxRead →