risk-based-testing

Quick Route

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it as the foundation and skip questions already answered there. Gather the rest from stakeholders across engineering, product, and operations.

Revenue-Critical Flows

• Which user flows directly generate revenue? (checkout, subscription, billing, upgrades)
• What is the revenue impact per hour of downtime for each flow?
• Are there time-sensitive flows? (flash sales, market-hours trading, payroll deadlines)
• Which flows have contractual SLAs with financial penalties?

Recent Failures

• What broke in the last 3 releases? What escaped to production?
• What were the root causes? (code defect, config error, third-party failure, data migration)
• What was the blast radius of each incident? (users affected, revenue lost, reputation impact)
• Were there near-misses caught late in testing that could have escaped?

Fragile Areas

• Which parts of the codebase change most frequently? (high churn = high risk)
• Which modules have the lowest test coverage today?
• Which areas have the most complex business logic or the most conditional branches?
• Which code was written by engineers who have since left the team?

Third-Party Dependencies

• Which external services does the product depend on? (payment processors, auth providers, CDNs, APIs)
• What is the historical reliability of each dependency?
• What happens when each dependency goes down? (graceful degradation or hard failure?)
• Are there single points of failure with no fallback?

Compliance and Data

• What regulatory requirements apply? (GDPR, PCI-DSS, HIPAA, SOC2, SOX, EU AI Act)
• What data is most sensitive? (PII, financial, health, credentials)
• What are the legal consequences of a data breach or compliance violation?
• Are there audit requirements that mandate specific testing evidence?

Core Principles

1. Not all features are equal. A bug in checkout that blocks purchases is categorically different from a misaligned icon on a settings page. Equal coverage everywhere wastes resources on low-risk areas while leaving critical paths under-tested.
2. Risk = Impact × Probability. Risk is not a gut feeling. It is a product of two dimensions scored independently: how bad if this fails (impact), and how likely to fail (probability). Score both consistently across the product, then multiply.
3. Risk assessment is continuous. A risk model created once and never updated creates false confidence. Risk changes when the product changes, when dependencies change, when the team changes, and after every production incident. Build reassessment into the rhythm.
4. Near-misses are data. A catastrophic bug caught in staging is not a success story — it is a signal that the model underestimated that area. Track near-misses with the same rigor as production incidents.
5. Risk informs coverage, not the other way around. Do not start with "we need 80% coverage everywhere." Start with "where would a failure hurt most?" and let the model drive coverage targets per module.

Workflow

1. Identify → 2. Classify → 3. Analyze → 4. Heatmap → 5. Coverage → 6. Reassess → (repeat)
                                 ↑
                       score ≥ 10 only; skip to 4 if no items reach threshold

Phase 1: Risk Identification

Enumerate everything that could go wrong. Cast a wide net. Sources include:

• Stakeholder interviews: Product managers know business-critical flows. Engineers know fragile code. Support knows recurring user complaints.
• Incident history: Past failures predict future failures. Review post-mortems from the last 6–12 months.
• Dependency mapping: List every external service, database, message queue, and third-party API. Each is a risk vector.
• Change analysis: Areas with frequent code changes have higher defect probability. Use the ranked churn command in Phase 6 to find them; git log --stat <file> is for per-commit inspection of a specific suspect, not for ranking.
• Architecture review: Shared databases, single points of failure, synchronous chains, and tightly coupled modules amplify blast radius.

Use HTSM v6.3 (Heuristic Test Strategy Model, Bach) as a Phase-1 lens — its state-based and boundary heuristics surface risks a pure feature-list misses. Download: https://www.satisfice.com/download/heuristic-test-strategy-model

Output: a raw list of risk items, each describing what could fail and what the consequence would be.

Phase 2: Risk Classification

Categorize each risk item along two axes.

Impact categories (how bad is it):

Probability categories (how likely is it):

Composite score = Impact × Probability. Frequent changes indicate defect probability, so a Moderate-impact (3) feature under heavy churn scores Probability 5 → Risk score: 15 → CRITICAL zone, despite "only" moderate impact. The composite score drives priority, not impact alone.

Phase 3: Failure Mode Analysis

For each high-risk item (score ≥ 10), perform a detailed failure mode analysis.

Feature/Component: [name]
Risk Score: [impact × probability]

Failure Mode 1: [what specifically can fail]
  Trigger:            [what causes this failure]
  Blast Radius:       [users affected, systems affected, data affected]
  Detection Method:   [how would we know -- monitoring, user report, test]

Example — E-commerce Checkout (Risk Score 20, Impact 5 × Probability 4):

AI/LLM failure classes

For AI/LLM features, classify against these CT-GenAI classes and score Impact and Probability independently like any other risk. The mitigation is the existence of an automated eval suite, not a single manual test.

AI/LLM-specific failure classes (from ISTQB CT-GenAI v1.1, effective 27 April 2026):

• Hallucination / reasoning error — Impact: moderate to major; Probability: high without explicit prompt-eval coverage. Detection: golden-dataset evals, fact-check assertions (see ai-system-testing).
• Bias — Impact: catastrophic in regulated industries (finance, healthcare, hiring). Probability: dataset-dependent. Detection: counterfactual evals, demographic-parity checks.
• Prompt injection / jailbreak — Impact: major (data exfiltration, prompt extraction). Probability: high for any externally-facing LLM feature. Detection: Garak, PyRIT, Promptfoo redteam.
• Privacy leak — Impact: catastrophic under GDPR/CCPA/EU AI Act. Probability: dataset-dependent. Detection: PII scanning of training data and prompts.
• AI Act / regulatory non-compliance — Impact: catastrophic (fines, ban). Probability: high for EU-facing AI features. Detection: see compliance-testing.

Tool freshness (mid-2026): PyRIT now lives at microsoft/PyRIT — the old Azure-hosted repo was archived March 2026, so do not point new redteam work at the legacy Azure path. Promptfoo was acquired by OpenAI (March 2026) but remains MIT-licensed. Garak is current and unchanged.

Reference frameworks: CT-GenAI v1.1 (ISTQB, effective 27 April 2026) codifies the AI/LLM classes above. WQR 2025-26 (Capgemini, 17th edition, Nov 2025) gives the adoption-stage framing for AI risk planning.

Phase 4: Risk Heatmap

Plot all risk items on a 5×5 matrix to communicate priorities and drive coverage decisions.

Zone boundaries and action mapping:

Populated example (where each named risk lands):

Phase 5: Test Coverage Alignment

Map test density to risk level. Every zone gets a prescribed approach.

Gap Analysis Worksheet

Compare current coverage against required coverage per risk zone:

A churn signal forces this worksheet open: a module that changed 47 times in 3 months (Probability → 5) with only 40% branch coverage and no integration tests jumps zones (e.g. MEDIUM → HIGH), and the new coverage target is justified by the churn, not picked arbitrarily.

See references/examples.md for four fully-scored examples (checkout, media platform, third-party API, auth) showing the path from risk score to prescribed coverage.

Phase 6: Monitoring and Reassessment

Risk assessment is not a one-time activity. Build reassessment into the team's rhythm.

Reassessment triggers:

• After every production incident (within 48 hours): re-score the affected items, check dependency health, and re-run the Phase-5 gap analysis to expose any coverage gap the incident revealed
• When a new feature area is introduced
• When a critical dependency changes (API version, provider switch)
• When team composition changes significantly
• Quarterly at minimum, even without triggers

Continuous risk signals to monitor:

• Code churn by module (ranked frequency table):

  git log --since="3 months ago" --name-only --format= | grep -v '^$' | sort | uniq -c | sort -rn | head -20
• Defect clustering: Which modules produce the most bugs? Track with issue labels.
• Near-miss frequency: How often do staging/QA catches prevent production incidents?
• Dependency health: Monitor status pages and uptime of critical third-party services.
• Coverage trends: Is coverage increasing or decreasing in high-risk areas?

Anti-Patterns

Testing everything equally

Applying the same coverage target to every feature regardless of risk. A 90% target on a settings page wastes effort that should go to payments or auth. Let the risk model drive allocation.

One-time risk assessment

Creating a matrix during planning and never updating it. The product, team, and dependencies all change. A model from 6 months ago is outdated and out of date the moment a dependency, feature, or incident shifts the picture — it does not reflect today. Schedule reassessment and enforce it.

Ignoring near-misses

Treating bugs caught in staging as pure successes. If a critical bug was caught only by manual testing, the automated safety net has a gap. Document near-misses and adjust the model.

Risk theater

Going through the motions (filling matrices, drawing heatmaps) without changing test allocation. If the heatmap exists but coverage does not align to it, the exercise was wasted. Verify alignment quarterly. Bolton's "Quality Engineering Is Not Testing" (2026-04-20) warns of exactly this — building a heatmap and calling it "QE done." Reference: https://developsense.com/blog/2026/04/quality-engineering-is-not-testing

Anchoring on historical risk

Over-weighting past incidents and under-weighting new vectors. A module that failed 2 years ago and was since rewritten may no longer be high risk; a brand-new third-party integration has unknown risk that deserves attention.

Confusing severity with priority

Severity is how bad a failure is; priority is how urgently to test it. A catastrophic-but-rare failure (earthquake destroys data center) can be lower priority than a moderate-but-frequent one (search occasionally wrong). Use the composite score, not impact alone.

Verification

Prove the matrix is real and aligned before calling it done — smallest check first:

• Artifact exists and is tracked: git ls-files | grep -E 'risk-matrix|qa-project-context' returns the file. An untracked draft on someone's laptop is not a risk model.
• Every in-scope feature is scored: grep the artifact for rows missing an impact or probability number. A feature with a name but no score is a gap, not a low risk.
• Top items have failure mode analysis: every item scoring ≥ 10 has a block with all five fields (Trigger, Blast Radius, Detection Method, Current Mitigation, Gap). A failure mode with an empty Gap line was not actually analyzed.
• Coverage aligns to the heatmap: for each CRITICAL/HIGH feature, confirm the prescribed coverage from the Phase-5 table actually exists (run the suite, check the coverage report's per-module numbers against the target). If the heatmap says CRITICAL but the module has 40% branch coverage and no E2E, the exercise was Risk Theater.
• Churn signal is current: re-run the Phase-6 churn command; any module in the top 10 that is not scored ≥ Probability 4 is a model that has drifted from reality.

Done When

• A scored risk matrix exists as a tracked artifact (.agents/qa-project-context.md risk section or a committed risk-matrix.md), with every in-scope feature scored on impact (1-5) and probability (1-5)
• Each feature's composite score places it in a named zone (CRITICAL, HIGH, MEDIUM, or LOW) with a corresponding testing action assigned
• Every feature scoring ≥ 10 has a completed failure mode analysis with trigger, blast radius, detection method, current mitigation, and gap documented
• Coverage requirements per zone are mapped against current coverage, with each gap assigned a Priority (P0-P3), an Owner, and a Target Sprint
• Reassessment triggers and cadence are recorded in the same artifact (quarterly minimum, plus within 48 hours of any production incident)

Reference Files (in references/)

• examples.md — Four fully-scored worked examples (checkout, media platform, third-party API, auth): risk profile → failure modes → prescribed coverage.

Related Skills

• test-strategy — The multi-quarter QA strategy this risk matrix feeds into; risk assessment is one input to a broader strategy.
• test-planning — Single-sprint/release planning uses risk priorities to decide what to test this iteration.
• release-readiness — Go/no-go decisions reference the heatmap to confirm critical areas are covered.
• qa-metrics — Defect escape rate and defect clustering feed back into risk reassessment.
• ai-system-testing — Building the eval suites that mitigate the Phase-3 AI/LLM failure classes.
• qa-project-context — Captures the critical flows, fragile areas, and dependencies this skill consumes and writes the risk matrix back into.