At a glance
- Owner
- StratQuad
- Version
- November 2026
- Applies to
- Organisations commissioning or running security testing
- Mandatory
- Guidance; testing is required by some regulators, contracts and schemes
Key points
- A penetration test is an authorised, scoped attempt to exploit weaknesses and show their impact. Authorisation is what makes it lawful under the Computer Misuse Act.
- Scope, rules of engagement and a named emergency contact must be agreed in writing before testing starts.
- In the UK, NCSC CHECK is the assured scheme for central government, public sector and critical national infrastructure systems; CREST accreditation is a widely used industry benchmark.
- Value comes from retesting and remediation, not from the report itself.
- Agentic AI offensive frameworks such as ARTEX and CyberStrikeAI, published as testing tools, were reported in use by threat actors against organisations in South Korea, Japan and elsewhere in 2026.
Types of test
| Type | Answers |
|---|---|
| External infrastructure | What can an internet based attacker reach and exploit? |
| Internal infrastructure | What can an attacker with a foothold, or a malicious insider, achieve? |
| Web application and API | Can users access data or functions they should not? |
| Cloud configuration | Are identities, storage and network controls configured securely? |
| Social engineering | Do people and processes resist phishing and pretexting? |
| Red team | Can a realistic adversary achieve a defined objective without being detected? |
Scoping and authorisation
- Objectives
What the test should prove or disprove, tied to a business risk.
- Scope
IP ranges, hostnames, applications, accounts and environments in scope, and anything explicitly excluded.
- Third parties
Written permission from cloud, hosting or managed service providers where their terms require it.
- Rules of engagement
Permitted techniques, testing windows, data handling, and whether denial of service or social engineering is allowed.
- Authorisation letter
Signed by someone with authority over every system in scope, naming the testing company and testers.
- Emergency contact and stop procedure
On both sides, for the whole testing window.
UK assurance schemes
- CHECK (NCSC): for testing central government, public sector and critical national infrastructure systems. An NCSC panel admits companies; CHECK Team Leaders and Members must hold UK Cyber Security Council Security Testing titles and be eligible for SC clearance.
- CREST: industry accreditation for testing companies, with examined individual qualifications.
- CBEST (Bank of England) and STAR-FS (Bank of England, PRA and FCA, with CREST accredited providers): intelligence led testing for regulated financial firms. STAR-FS is firm led and needs less regulatory effort than CBEST.
Getting value from a test
A good report gives an executive summary in business terms, findings with evidence and reproducible steps, a severity rating with its method stated (usually CVSS), and specific remediation. Book the retest when the test is booked, and track findings in the same system as other vulnerabilities.
Evidence expected
Keep: the scope and rules of engagement, the signed authorisation, the report, the remediation plan with owners and dates, and the retest results. Auditors for ISO/IEC 27001, CAF and DSPT ask for this chain.
AI agent driven testing
Testing firms increasingly use large language model agents to run reconnaissance and exploitation at scale. The same tooling is available to attackers. ARTEX, a multi agent testing framework released as open source, and CyberStrikeAI, an open source framework with coordinated agents and over 100 tool recipes, were both reported in attacks in 2026: ARTEX against South Korean financial institutions and Japanese companies, and in sign in attempts on Microsoft 365 tenants through compromised application identities; CyberStrikeAI against organisations mostly in Japan. CyberStrikeAI includes approval gates and audit logs, but the Cloud Security Alliance notes that nothing in its design stops an operator who controls the deployment from disabling them.
Practitioner note
Two practical consequences. When buying a test, ask whether AI agents will be used, which model providers will see your data, and how agent actions are constrained and logged. When defending, expect distributed, low rate attempts across many assets rather than noisy bursts: threshold based alerts tuned for a human attacker can miss an agent that throttles itself per target. Continuous external attack surface testing has become a baseline control.
Common mistakes
- Authorisation signed by someone who does not control all systems in scope.
- Scope limited to what is known to be secure, so the test proves little.
- Findings accepted without a retest, or left in a PDF rather than the remediation tracker.
- Annual testing treated as the only assurance, leaving months of exposure between tests.