Mobile app security testing: what to expect and what to ask for
Compare mobile security testing approaches and agree scope, evidence, reporting and retesting before commissioning an assessment.
If you commission or buy a mobile app, someone may tell you it has been "security tested". That phrase covers very different work. OWASP's Mobile Application Security Testing Guide (MASTG) has a chapter that explains the main kinds of testing, how a penetration test is usually organised and how security fits into a development process. We read that chapter in full on 15 November 2026. We did not read the guide's individual test cases or the MASVS profile definitions. Everything attributed is OWASP's. This post is a buyer's guide to the terms, not a testing procedure, and it does not make any app safe.
Three ways a tester can start
- Black-box: the tester knows nothing about the app, like a real outsider.
- White-box: the tester has source code, documents and diagrams, so testing is faster and can be more detailed.
- Gray-box: some information is given, usually credentials. OWASP says this is the most common kind in the industry.
The guide advises requesting the source code so testing time is used efficiently, and calls a white-box test the way to go if the app has not been tested before. Code access does not simulate an outside attack, but it lets the tester check every suspicious behaviour in the code. (Its other reason: Android code can be decompiled but may be obfuscated, and de-obfuscating takes time.)
Static and dynamic analysis
- Static analysis (SAST) examines the app's code without running it, manually or with tools. Manual review is good at business-logic flaws, design flaws and standard violations in code that is technically secure but logically wrong, which automated tools are unlikely to find. It needs an expert who knows the language and frameworks and can be slow on big code bases. Automated tools check against rules but may produce many false positives unless configured for the target, so a security professional must review results.
- Dynamic analysis (DAST) examines the running app, both on the device and against the backend APIs, and is usually used to check protection against common attacks: data exposed in transit, authentication and authorization problems and server configuration errors.
Why scanner output needs judgement
The guide gives an example. Scanners built for websites report CSRF and cross-site scripting that may not matter for a mobile app. CSRF needs the user's browser to add a session cookie automatically to a request triggered by a malicious link; the guide gives the default-browser/separate-cookie-store flow as an example. Do not assume every app, WebView or deep-link flow behaves that way: check actual navigation, credential attachment and backend exposure. Stored XSS can matter if the app uses WebViews, and more so if it exposes JavaScript interfaces; reflected XSS is rarely an issue. Its advice: consider exploit scenarios in the risk assessment and do not blindly trust a scanner's output. A similar point applies to code: using a weak random number generator to shuffle a game list is not a security flaw, but using one to create keys is; and Base64 is encoding, not encryption.
How a penetration test is usually structured
OWASP lists five phases:
- Preparation: scope, which security controls apply, testing goals and sensitive data, and legal protection of the tester. The guide says attacking a system without written authorisation is illegal in many parts of the world. Decide the security level before testing; the guide describes MAS-L1 tests as applicable to all mobile apps and says walking through MAS-L1 and MAS-L2 with technical and business stakeholders is a good way to choose coverage. It notes regulations may make some tests relevant, giving the example of two-factor authentication being obligatory for a financial app in some countries.
- Intelligence gathering: environmental information (the organisation's goals, industry, stakeholders, internal processes) and architectural information (the app, operating systems, network and remote services).
- Mapping the application: entry points, features and data, ranked by potential damage, with test cases. Threat modelling documents help a lot, and the guide says threat modelling should be part of the development life cycle and usually occurs early.
- Exploitation: confirming which findings are real. OWASP lists five axes for judging relevance: damage potential, reproducibility, exploitability, affected users and discoverability. It admits time or budget often limits tests to mapping with automated scanners.
- Reporting: a useful report has an executive summary, scope and context, methods, information sources, prioritised findings, detailed findings and recommendations to fix each.
Things to settle with the tester before they start
- A definition of sensitive data. The guide says detecting leakage may be impossible without one, and gives a default list when there is no policy: credentials and PINs, personally identifiable information that can be abused for identity theft, device identifiers that may identify a person, highly sensitive data, data protected by law, and technical data used to protect other data, such as encryption keys. It also describes data at rest, in use and in transit.
- Two builds. Controls such as certificate pinning and root detection slow testing. The guide suggests asking for a release build to confirm the controls work and a debug build with some controls off to cover the other test cases (not always possible, depending on scope).
- Test environment. Company networks and policies may block rooted phones or network tools.
Testing during development, not only at the end
The guide says a "classical" test near the end finds issues late, and that in a waterfall process changing architecture after the testing phase is hard. For Agile and DevOps work, it describes automating security activities across the life cycle: static analysis in continuous integration, dynamic analysis after the build and before delivery, manual checkpoints between phases, regular scanning in operations, pentests of the production version in a dedicated environment with production-like data, and active monitoring. It says the commit stage should usually finish within five minutes and not more than ten, which many security tools cannot meet, so heavier checks run later. It also gives a lifecycle: risk assessment, security requirements (with the MASVS and MASWE as a source), threat modelling, secure development with code review and static analysis, release-candidate pentesting, acceptance, operation and decommissioning.
A test strategy does not need to be a formal document, but it must be shared and agreed by the technical teams. The guide gives example metrics (security stories completed, unit-test coverage of security controls, security bugs per build and the trend of the security bug backlog) and says internal tests suit daily improvement while independent third-party tests benefit the organisation as a whole and are required by many regulations. The guide describes periodic third-party tests as a common arrangement, not a maximum safe frequency. Set testing and retesting frequency from risk, releases, incidents and applicable obligations.
Questions to ask a vendor
- Black-box, gray-box or white-box, and did you have our source code?
- Which MASVS controls and profile were in scope, and which were excluded and why?
- What was tested statically and what dynamically, and were scanner findings checked by a person?
- Was it the release build, and what was disabled in the debug build?
- Can we see the report, with scope, methods, prioritised findings and fixes, and a retest after fixes?
- Is testing built into your pipeline or done once at the end?
Limits
One MASTG chapter, read in full. The MASTG's actual test cases and techniques, the MASVS profiles and any tool were not read, so we cannot say how any control is tested in detail. Legal requirements for testing and for regulated apps (finance, health) differ by country; take advice from counsel. We name no testing vendors or tools. No statistics.
By Omnistu.