A Free Attack Tool Cuts a Break-In to Three Hours

The number in the FBI’s bulletin is three hours. That is how long the bureau says it took, in several recent intrusions, to go from the first scan of a target network to a working set of credentials and a foothold inside it.

The tool the attackers used is called Artex, and it is open source. The FBI described it in a bulletin dated Nov 16 as an artificial intelligence system that strings together the steps an intrusion normally requires: scanning for exposed services, guessing and testing credentials, and moving from one machine to the next once access is established. It is distributed through public repositories, where it keeps getting updated.

Three hours is not a remarkable figure for a skilled human operator with a specific target. It is a remarkable figure for a general-purpose tool that anyone can download, because the expensive part of an intrusion has never been the exploitation. It is the reconnaissance and the patience: mapping a network, identifying which of a thousand exposed services is worth attacking, working out which account has the permissions that matter, and doing all of it without tripping a detection.

Automation collapses that work. A language model reading scan output and error messages can hold a picture of a network in a way that a script cannot, and it can adjust when the first approach fails. That is the difference between a scanner and an operator. Scanners have been free for twenty years. Operators have not been.

The distribution model matters as much as the capability. Commercial attack tooling arrives with a license, a vendor and a support contract, and a vendor can be pressured, sued or sanctioned. An open-source project has none of those handles. There is no company to notify, no patch to withhold, and no version to recall. The code can be forked, renamed and improved by anyone who finds it, and the improvements are public by design.

For defenders, that reverses a customary asymmetry. Security teams have long relied on the attacker’s need for expertise, and on the cost of acquiring it, as an implicit control. A tool that supplies judgment at no cost removes that control without adding a single new vulnerability. The software being attacked is the same software it was last year.

The bureau’s recommendations are modest. It advised organizations to require multi-factor authentication on internet-facing edge devices, and to restrict outbound network access for automated tools. Both are sensible, and both are floors rather than answers.

Multi-factor authentication on a firewall, a VPN concentrator or a remote access gateway stops the part of Artex’s workflow that tests stolen passwords. It does not stop the tool from finding a device without that protection, and a large organization will have one. It also does not address session theft, where an attacker takes a valid token rather than a password and never has to authenticate at all.

Restricting outbound traffic is the harder instruction. Automation needs egress to reach repository mirrors, package indexes and model endpoints, and an attacker running inside a corporate network needs exactly the same access. A list of permitted destinations is straightforward to write and difficult to keep current, and every operations team that has tried to lock down egress knows the shape of the argument that follows: the build pipeline breaks first, and the exception is granted before the analysis is finished.

What the bulletin describes, then, is a shift in the tempo of intrusion rather than in its methods. The vulnerabilities being used are ordinary. The targets, internet-facing appliances at the edge of corporate networks, are the ones that have been attacked for a decade. What has changed is how fast the sequence runs, and speed is what detection systems are worst at.

Most enterprise defenses are built on the assumption that an attacker needs time. A signature catches known tools. An anomaly detector catches behavior that departs from a baseline over hours. A human analyst catches what the first two miss, provided there is enough of a window to look. Compress the sequence into a single afternoon and each of those layers gets less to work with, and the response has to start before the scope of the incident is known.

There is a policy dimension the bulletin does not address. Export controls and licensing regimes are built to restrict access to models and to the hardware that trains them. They do not reach a piece of open code that calls a model through a public interface. The tool is not a weapon that can be scheduled. It is a script anyone can run against a reachable target. Regulators have spent two years arguing about who may train what. The document published this week is about what can be done with the results.

For an individual company, the practical response is a measurement rather than a purchase. The relevant question is not whether the network can be defended against automation, because it cannot be defended against a determined operator with or without it. The question is how long the organization would take to notice, and how much of that time is spent waiting for a person to look at an alert that a machine already ranked as urgent.

Three hours is a useful benchmark for that exercise, if only because it is short enough to be uncomfortable. A security team that models its own detection timeline against it will usually find that the detection is fast and the decision is slow. The gap between the two is where the next incident will happen, and no bulletin can close it.

The FBI said it would continue to track the tool and related threats. It did not say how many organizations were affected or whether any data was taken. Those are the details that would tell a defender whether Artex describes the present or predicts the next year.

Related Posts

  • November 20, 2026
  • 17 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 15 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…