CYBER ARMOUR Scoping call
Approach

How an engagement runs

Seven phases, each with a named owner. Where we use automation, the phase says so.

01Method

Seven phases, one owner each

An engagement is not one activity, and the split between people and tooling is different in each phase. This is how the work is actually divided.

01

Scope and threat model

We start from what would hurt you and work backwards to the attack surface.

Operator
02

Reconnaissance and attack surface mapping

Full enumeration: hosts, services, subdomains, cloud assets, exposed credentials, public code. Automation runs the sweep. Operators set the hypotheses it is sweeping for.

Operator + tooling
03

Exploitation

Every action against your systems is chosen and executed by an operator, and we can tell you why each one was taken.

Operator
04

Variant analysis

Each confirmed bug becomes a search pattern applied across the whole estate. Where a pattern returns many hits we verify a representative sample by hand and report the remainder as a class, labelled as such, with the verified samples named. We never fold unverified hits into the verified count.

Operator + tooling
05

Chaining and impact

Individual findings assembled into the paths that matter: a forgotten staging host to domain admin, a read-only API key to your customer database.

Operator
06

Verification

Every finding reproduced by hand with a proof of concept you can run yourself. Nothing ships on the word of a tool.

Operator, mandatory
07

Reporting and retest

Written for two audiences at once: the engineers who have to fix it and the executives who have to fund it. Retest of remediated findings is included in the fee.

Operator
02On AI

Where AI fits, and where it does not

Every security vendor now claims AI, and most of them mean a scanner with a chat box bolted to the side. We use these models daily. They have changed how much ground we can cover. They have not changed who does the difficult part. We ran them inside our own shop for eighteen months before they touched a client engagement, and everything below happens on our own cluster — none of it involves sending your data anywhere. Here is our position.

Where it helps
  • Breadth, at speedTen thousand hosts, every subdomain, every parameter on every form, enumerated and triaged in hours rather than weeks. More of your estate gets tested.
  • Reading code quicklyTracing a tainted value from HTTP request to SQL sink across forty files, for every entry point rather than the dozen there was time for.
  • Variant huntingOnce an operator confirms a bug, sweeping the estate for every other instance of the same pattern. One finding becomes a class of findings.
  • Tooling on demandA protocol parser, a targeted fuzzer, a harness, so a speculative idea costs an hour instead of a day.
  • Correlation across sourcesSource, cloud configuration, directory structure and logs read side by side, so the relationships between them surface faster.
Where it does not
  • Business logicA model does not know that your approval workflow is where the money is, that this role was never meant to see that account, or that a discount code is a financial control. Abuse of intended functionality has no signature to match against.
  • ChainingReal compromise is rarely one critical. It is five findings rated low, composed in the right order. Tools score findings one at a time. Attackers think in paths.
  • Novel exploitationWriting a working exploit for a bug nobody has published, against modern mitigations, on a target that fights back. Models converge on the most probable answer, and the attacks that hurt you are the improbable ones.
  • Knowing when to stopWhether an action is safe against production at 14:00 on a trading day is a judgement call with consequences attached. It is not a parameter.
  • Confident inventionAI produces vulnerabilities that do not exist, described fluently and scored precisely. An unverified finding is not a finding. It is a rumour with a CVSS number attached.

So we use it the way we use any other tool: for coverage, never for conclusions.

The machine proposes, an operator proves. Ask any firm which parts of your report a person verified. The answer should be all of them.

03Handling

Where your engagement data goes

If a firm is going to point tooling at your source code and your network, you are entitled to know exactly where that data goes. These are the commitments we work under.

  1. 01
    Your data never leaves our infrastructureThe models we use run on a dedicated NVIDIA DGX cluster we own and operate. No client source code, credentials, configuration or findings are sent to a third-party AI service. There is no provider to name, because there is no provider — and nothing to train on, because your data never reaches anyone else's model.
  2. 02
    Custody and destructionEvidence, findings and reports stay on infrastructure we control, encrypted in transit and at rest, and are destroyed on the schedule your contract specifies.
  3. 03
    An auditable trailOn request we provide a timestamped log of the activity run against your estate, automated and manual, including the source addresses we tested from.
  4. 04
    Rules of engagement in writingScope, timing, escalation contacts and stop conditions agreed before anything starts, and respected once it does.
—Contact

Tell us what you are worried about

One call is enough to scope an engagement. We will tell you whether we are the right firm for it, and what it will cost.

[email protected]
  • Target

    What should be tested: an application, a network, an Active Directory estate, a cloud tenant, an AI system, your people, or all of it.

  • Concern

    What you are worried about: ransomware, a regulator, a customer security questionnaire, a breach you already had.

  • Timing

    Any deadline you are working against: an audit date, a release, a board meeting, a contract renewal.

  • Constraints

    Production systems that cannot be touched, change freezes, legacy you already know about, or testing you have had before.