18 November 2026

If attackers use AI faster than you, you lose

If attackers use AI faster than you, you lose
4 min. readBy Borg

On Friday 13 November the FBI sent out a warning about a hacking tool called ARTEX. It is an open-source framework that uses AI agents to do what a penetration tester does: map a target, find the weak points, exploit them and move deeper. According to BleepingComputer's report of the advisory, several groups have been using it against real organisations since at least August.

Two sentences in the advisory matter more than the rest. The first says that some intrusions are "progressing from initial targeting to privileged access in under three hours". The second says the framework "enables threat actors with minimal technical expertise to conduct complex, multi-target operations simultaneously".

The first is about speed. The second is about who can now do this, and to how many targets at once. This post is about what that changes for the people defending, and why testing once a year stops being enough.

What changed in a year

Twelve months ago this was news of a different kind. In November 2025 Anthropic reported the first AI-orchestrated espionage campaign it had seen. A state-sponsored group had used an AI coding tool to carry out 80 to 90% of an operation against roughly thirty targets, with people stepping in at a handful of decision points.

That took a well-resourced group. What the FBI describes now is a tool anyone can download. Its developer took the original repository down in October, and the FBI says forks and archived copies are still easy to get. The skill has moved from the person into the software.

Speed is half of it, scale is the other half

A skilled attacker works on one target at a time, and there are not many skilled attackers. That has always been a quiet part of every organisation's defence. Most companies were not worth someone's week.

An agent has no week to protect. It runs all day, against many targets at once. The same BleepingComputer report cites CrowdStrike on one group that pointed ARTEX at more than 200 organisations in 14 countries. At least nine were compromised. Most attempts failed, and when trying costs almost nothing, they are allowed to.

What it finds is ordinary

None of this depends on a new kind of vulnerability. The FBI says the tool is particularly effective at finding weaknesses in overlooked systems: partner portals, employee applications, legacy services and third-party integrations. TechCrunch's account describes the way in as simple security bugs, weak protections and exposed or stolen passwords.

These are the findings that fill pentest reports every year. The forgotten subdomain. The service nobody owns. The credential that was never rotated. What has changed is how long they can sit there before someone tries them.

Why a yearly pentest falls behind

A pentest is a snapshot. The report describes your systems as they were in the weeks the test ran. By the time it has been read, prioritised and worked through, it describes a version of the company that has shipped many times since.

That gap was tolerable when the people looking for the same holes were slow and few. It is not when the looking is automated. A service deployed in March and tested in November has spent eight months in front of tools that need an afternoon.

Attack your own systems first

The answer follows from the problem. If the other side tests your systems all the time, the only way to see what they see is to do the same: from the outside, with the same kind of tooling, before they do.

In practice that means three things. Test when something changes, not when the calendar says so. Count a finding as real only when it comes with proof that it worked. And get it to the person who can fix it while the change is still fresh.

This is what we build at Borg. Our red team agents attack an organisation's own systems the way an outside attacker would, keep doing it as those systems change, and report what they actually got into, with the evidence.

Where this stops

It does not replace people. An agent is good at breadth: every host, every endpoint, every release. A human tester is still better where the flaw is in how the business works, such as a refund flow that can be abused. Keep the yearly test for depth, and stop asking it to provide coverage.

Finding is also not fixing. A faster stream of findings only helps if someone owns them. And AI-driven attacks are not invisible: the FBI notes that they are noisy and can trigger alerts, so monitoring still matters. It cannot be the whole plan when the window is three hours.

Borg on
Social Media

Borg specializes in securing complex, high-risk digital systems.