The project ended. The system stayed online.

Digital transformation adds systems faster than a company lists them, and whatever nobody listed is now the first place an automated attack begins.
Eighteen months ago the sales team trialled a portal for distributors. The pilot ran for three months, the result was lukewarm, and the company chose something else. The project team moved on, the vendor sent a final invoice, the steering committee closed the item. The portal is still running. It has its own address, its own login screen, and a test account whose password six people knew, two of whom have since left. Nobody patches it, because nobody remembers it exists.
No single person was careless here. This is what a project does when it has a start date and no switch-off date.
What changed: looking became cheap
Until recently, a portal like that was protected by the fact that finding it was not worth anyone’s time. Mapping someone else’s infrastructure was human work, and humans picked targets that repaid the effort.
In 2026 came ARTEX: an open-source system for autonomous penetration testing in which several agents built on language models share the work. One breaks the goal into steps, one plans, several more each carry out a single task with real tools, and everything they find goes into a shared graph of what exists and what has already been tested. It was built for testing that the system’s owner has authorised. In October 2026, however, reports described a campaign in which an unidentified operator turned it on financial institutions in South Korea; the Cloud Security Alliance research note collects what is known. Investigators stressed that the tool did not act alone. A person was running it.
ARTEX originated in the Autumn-27/ARTEX repository on GitHub; many forks circulate today. Defensive material has been published alongside one of those forks: a description of how to recognise such an attack and how to harden against it. For anyone leading a transformation, one paragraph in it matters most. The authors write that such an agent’s first target is not a clever unknown flaw but a known vulnerability left unpatched, and that the way in is often an auxiliary system: one for partners, for recruitment, for employees. Not the main service, but whatever stands beside it.
The agent does not choose. It lists everything a company has exposed and works through it in order, without tiring and without breaks. A portal left over from a pilot is as good a starting point as the transaction system, only less watched.
Where auxiliary systems come from
A pilot with no end date. The pilot plan says when the environment is created and what it has to prove. It rarely says who switches it off, and when, if the result is negative or the company picks something else. We wrote about a successful pilot being the start of a different project in the piece on scaling after a pilot. A failed pilot has a sequel too. Nobody plans that one.
A portal that outlived its project. An event site, an application form for one hiring campaign, a supplier site from the days of the previous purchasing system. Each had an owner while the project lasted. Once the budget closed, the owner became “IT”, which is to say nobody in particular.
An integration built as a trial. To see whether two systems could be connected, someone exposed an interface with an access key. The trial worked, the real connection was built differently, and the old interface and its key stayed.
A tool started outside any project. A team set up a board, a database or an agent by itself, because that was faster. That is rational behaviour, as we argued in the piece on people who already use AI. The trouble starts when the tool gets a public address and access to data, and never reaches any list.
Access for a vendor who has finished. An implementer’s account, remote service access, a test environment kept “just in case”. The contract expired. The permissions did not.
The cause underneath all five is the same: companies hold projects to account for what they launched, not for what they left behind.
Security as a line in the plan, not a separate programme
The defensive material published alongside one of the forks leads to a list of technical recommendations: an inventory of what is exposed to the internet, fast patching, a second login factor, permission checks on the server side, network segmentation. That is work for security people, and VUCALE does not do it. We do not run penetration tests, monitor networks or deploy attack-detection tooling.
Part of that list, though, can be dealt with earlier and more cheaply, in the way change itself is run.
Every system has a named owner. Not a department, not a vendor, not a project manager whose role ends with the project. A person who answers for the system being patched, and who has the authority to switch it off.
Every pilot gets its switch-off date on the day it starts. On the pilot charter, next to the success criterion, sits a date after which the environment disappears unless someone decides to keep it. Keeping it takes a decision; switching off happens by itself. In most companies today it is the other way round.
Closing a project includes cleaning up. A list of the accounts, keys, addresses and vendor access the project created, with a note against each: handed to an owner, or removed. The steering committee does not close the project until that list is empty. It is one of the few decisions a committee can genuinely take, as we wrote in the piece on the steering committee that steers nothing.
The inventory comes before the next rollout. Before the company launches the next portal or the next agent, it checks what is still running from earlier initiatives. A new system added to an unknown set of old ones enlarges a problem nobody can see.
Auxiliary systems are held to the standard of the main ones. The partner portal and the candidate form get the same login and the same updates as the system the company lives on. If that is too expensive, it is a sign the system should be switched off, not kept in a worse state.
None of these lines needs a new tool. Each needs someone to add one column to the transformation plan: who switches this off, and when.
What to measure
| Vanity metric | Outcome metric |
|---|---|
| number of pilots launched | number of pilots switched off by the date written down at the start |
| “we have a security policy” | how many systems on a public address have a named owner |
| number of integrations delivered | how many vendor keys and accounts were removed when recent projects closed |
| share of projects closed on budget | days between a project ending and its test environments being switched off |
A test for today
Write down the projects and pilots of the last two years. A list from memory will do, ten or fifteen items. Against each one, ask a single question: what from this project is still running on a public address?
Then add two columns: who owns it, and when it was last updated. If nobody at the table can answer for an item, that does not mean the answer is “nothing”. It means the only one who knows the answer today is whoever is looking at your addresses from the outside.
Pass the items without an owner to whoever is responsible for security this week. And add a switch-off date to your project charter template before the next pilot starts. If you want to talk about arranging a change plan so that it cleans up after itself, get in touch.
- security
- pilots
- transformation plan
- AI agents
- ARTEX