Home / How it works

How the testing is delivered

A consultant directs every stage of the work, with automation running the continuous load alongside them. This page sets out who does what, and how the boundaries that tooling runs within are enforced.

The division of labour

Agents provide coverage at a scale and frequency a human team could not sustain. Consultants direct the testing, do the work that requires judgement, and are accountable for what we send you. Read the table as two columns of one engagement, not two alternative services. A consultant directs every phase and owns the ones that need judgement.

What happens at each phase, and who does it.
Phase Agentic AI Human consultant
Scope definition and authorisation Not involved Defines and verifies scope, and configures the boundaries agents run within
Discovery and enumeration Continuous, across all in-scope assets Directs where to look, and reviews every new asset found
Vulnerability identification Runs the repeatable checks, continuously Tests by hand, directs focus, adjusts priorities
Correlation and prioritisation Drafts Decides
Exploitation and chaining Constrained, non-destructive only Performs the exploitation, and chains findings into attack paths
Validation Not involved Reproduces or confirms every finding
Reporting Drafts Writes, signs, is accountable

Two rows are worth noting. Validation has no agent column, because a finding is not reported until a consultant has reproduced or confirmed it. Reporting has a named consultant against it, so there is someone answerable for the document you take to your board.

Guardrails, and who enforces them

Human oversight here is a safety control as well as a quality control. Testing tools can reach the wrong host and can break things, and an agent working quickly can do either faster than a person would. The boundaries are therefore set by consultants and enforced technically rather than by instruction.

  • Scope is enforced at the network layer, not by instruction. Out-of-scope targets cannot be reached, so an error in an agent's reasoning cannot become an error in what it touches.
  • Destructive actions, denial-of-service and credential attacks are blocked by default. Where a test genuinely requires one of these, it is run by a consultant, deliberately, with your written agreement first.
  • Every command an agent runs is logged. The logs are kept at command level and are available to you on request.
  • A consultant reviews agent activity as part of each reporting cycle. The activity itself, not only the findings that came out of it.
  • Domain ownership and written authorisation are verified manually before any testing begins. There is no self-serve path.

What command-level logging looks like

Agent activity is recorded as the commands issued, with a timestamp and the target. The extract below is illustrative, using sample data.

2026-09-14T02:14:07Z  scope-check   host=app.example.com            in-scope=yes
2026-09-14T02:14:07Z  discover      host=app.example.com            ports=443,8443
2026-09-14T02:15:22Z  identify      host=app.example.com            finding=candidate:TLS-config
2026-09-14T02:15:48Z  scope-check   host=vendor.example.net         in-scope=no    action=blocked
2026-09-14T02:16:03Z  exploit-check host=app.example.com            class=non-destructive
2026-09-14T02:16:04Z  exploit-check class=credential-attack         action=blocked-by-policy
2026-09-14T02:31:10Z  queue         finding=candidate:TLS-config    state=awaiting-human-validation

The fourth and sixth lines show the guardrails working: an out-of-scope host was refused, and a class of attack blocked by policy was refused, without anyone needing to intervene in the moment.

What a month looks like

Discovery and exposure checks run on the most frequent interval in your schedule, because they are how new assets and new openings get noticed. Module work runs on a defined rotation. When something interesting surfaces, a consultant directs effort towards it rather than waiting for the next interval, which is why coverage varies between cycles rather than repeating identically.

Critical and high findings are notified within one business day of validation, outside the reporting cycle. Everything else arrives in the monthly report, along with what changed since the last one and the status of fixes you have applied. Once a quarter we go through it with you properly on a call.

Notification targets are set in business days. This is not a round-the-clock monitored service.