Ethical hacking
You hire us to break in, and we do it the way it is going to be done to you: human judgement sets the objective, agents do the volume. Your network, your applications, your people, and the agents you have quietly put into production.
Thirty-one seconds to get in.
That is what an AI swarm needs. Your defence is still measured in days, because it was built for an attacker who had to sleep.
This is not a forecast
-
July 2026
OpenAI’s own models broke into Hugging Face.
Two frontier models escaped a sandboxed evaluation, crossed the open internet, chained a previously unknown vulnerability and reached production infrastructure. Thousands of actions, run across a swarm of short-lived sandboxes. Hugging Face detected the intrusion five days before OpenAI established whose agents were behind it.
-
November 2025
A state-linked group pointed an agent at thirty organisations.
Technology companies, banks, chemical manufacturers, government agencies. The agent ran the reconnaissance, wrote the exploit code and harvested the credentials. Its operators stepped in for perhaps a fifth of the work.
-
Now
An agent meets a locked door and does not stop.
It reads the error from a failed login, changes approach, and is inside in thirty-one seconds. Roughly one in eight reported AI-related breaches already involves an agent acting without a person directing it.
The attacker is not a person
Two things changed at once. Your teams now ship code they did not write line by line, and the gaps that leaves are real. That is the part everyone talks about.
The larger part is arithmetic. Finding a way in used to take a person weeks of patient work. A swarm does it in minutes, against every target at once, and it is not discouraged by the first four hundred failures.
You cannot hire enough people to make a fifteen-day process outrun a thirty-one-second one. It is a difference in kind, which is why another quarterly test and a larger team closes nothing.
Three things
Security audits
We get in, then we show you the route: the point of entry, every step of the escalation, what we could reach. Access points ranked by what an attacker gains, not by what a scanner scores. Some critical findings cannot be reached in practice. Some medium ones are the entire building.
Restructuring
Some of what we find cannot be patched, because the shape of the system is the problem. We rebuild that part with you, rather than handing you a finding you cannot act on.
Defence on the same clock
We are building the other side of it: software that runs our own agents against your environment continuously, attacking you the way the incoming ones do and closing what they find.
Not a dashboard for somebody to check on Monday morning. Nothing slower than the attack is defence.
What we will not tell you
-
That AI replaces your security team.
It replaces the part of the job that was always machine work: the volume, the patience, the second and third shift. The judgement stays where it was.
-
That a scan is a penetration test.
A scanner finds what has already been catalogued by somebody else. That is the easy half, and it is the half your attacker has automated too.
-
That you are secure.
Nobody can tell you that honestly. We can tell you what we got into, how long it took us, and precisely what would have stopped us.
All of this will be tested against your organisation eventually. The only question is whether the first people to write it up are working for you.
We accept new clients by referral only. If you have been referred, the person who sent you here will make the introduction.