Back to Blog
Testing guide

What Is a Race Condition: Patterns, Examples, and Fixes

Race conditions are a hurdle in software testing. Learn what they are, the common patterns behind them, why they evade your pipeline, and how to fix them.

Armish Shah
August 7, 2026
September 4, 2026
What Is a Race Condition: Patterns, Examples, and Fixes

Testing guide

What Is a Race Condition: Patterns, Examples, and Fixes

by:

Armish Shah

September 4, 2026

8

min

Share:

What Is a Race Condition: Patterns, Examples, and Fixes | TestFiesta
On this page

Ready to take your testing to
the next level?

Sleek and intuitive workflows
Transparent pricing
Easy migration

Introduction

A race condition is a bug where the outcome of your code depends on timing you don’t control. In a race condition, two operations overlap, each one correct on its own, and together they corrupt the data, such as an inventory count, double-charging a customer, or handing an attacker root access. These outcomes can pass code review, survive 100% test coverage, and only show up under real concurrent traffic. This guide breaks down how race conditions work, why your existing pipeline can’t catch them, and how to fix them at the layer where they actually live.

What Is a Race Condition in Software?

A race condition occurs when a program’s behavior depends on the sequence or timing of events it doesn’t control, and at least one possible ordering produces a wrong result. The code assumes it’s the only thing running, which is false in the real world.

The classic example: two users see one item in stock. Both requests read stock = 1. Both pass the if stock > 0 check. Both decrement. Final stock: -1. Neither request saw the other’s write, because both reads happened before either write committed.

Nothing in that code is broken in isolation. Run it once, it works every time. Run two copies at the same moment, and it fails, not occasionally but reliably, whenever the timing lines up. That’s the defining trait of a race condition, correctness that depends on an ordering nobody guaranteed.

Race Condition vs. Data Race: A Distinction That Actually Matters

Most guides use these terms interchangeably, but they’re not the same thing, and the confusion causes real problems in code reviews and security triage.

Data race: A formally defined term. Two threads access the same memory location at the same time, at least one access is a write, and no synchronization sits between them. The C11 and C++11 memory models define this as undefined behavior. Data races are mechanical enough that tools can catch them: ThreadSanitizer and Go's -race flag detect them reliably at runtime.

Race condition: A semantic error. The program produces the wrong result because of the timing or ordering of events, whether or not a data race is present. No tool can detect this class in general, because detecting it requires knowing what the code is supposed to do.

So when a tool reports “no data races found,” that is not a clean bill of health. It means one specific, narrow class of concurrency bug is absent. The inventory oversell above can happen in code with no data races at all, because the race window sits in the database, not in memory.

The Two Race Condition Patterns Behind Most Production Incidents

Most production incidents caused by concurrency stem from two primary race condition patterns:

1. Check-Then-Act

The pattern: read a value, make a decision based on it, then act, assuming the value hasn’t changed between the read and the action. In a concurrent system, that assumption fails whenever there’s a gap between the check and the act. And there’s always a gap.

Three scenarios that show up in incident reports constantly:

  • Inventory oversell. Read stock = 1, gap, decrement. Two concurrent requests both read 1, both pass the check, both decrement. Final stock: -1. The fulfillment team ships an order that can't be filled.
  • Coupon abuse. Read coupon_used = false, gap, mark used. A user fires 50 simultaneous requests at the redemption endpoint. 47 of them pass the check before any write commits. One promo code, applied 47 times. Finance notices at month-end close.
  • Double-spend. Read balance = $100, gap, deduct $100. Two simultaneous transfer requests both see $100, both pass, both deduct. $200 leaves a $100 account. Discovered in reconciliation, not prevented at the source.

The rule worth internalizing: Any SELECT followed by a conditional UPDATE in separate statements is a check-then-act. In a concurrent system, it is vulnerable by construction. 

2. Read-Modify-Write

Read-Modify-Write (RMW) is a sequence of three operations performed on shared data:

  1. Read the current value from memory.
  2. Modify that value (e.g., increment, decrement, update).
  3. Write the new value back to memory.

The problem is that these three steps are not atomic (they don’t happen as a single indivisible operation). If multiple threads execute them simultaneously, a race condition can occur. 

An example:

Incrementing a Counter: Suppose two threads share a variable counter = 5. Both threads execute counter = counter + 1;. Internally, this becomes:

Step Thread A Thread B
Read Reads 5 Reads 5
Modify Calculates 6 Calculates 6
Write Writes 6 Writes 6

Expected result: 7

Actual result: 6

One increment is lost because both threads read the same original value before either wrote back the update. This is called a lost update, one of the most common race conditions.

Why Race Conditions Are Not Detected in Your Pipeline

Race conditions slip through the cracks because every standard quality gate tests a dimension these bugs don’t live in.

They're nondeterministic by nature. The same code path produces different results depending on thread scheduling, which the OS controls, not you. The bug disappears when you rerun the test. It disappears when you add a log line, because logging adds latency and latency changes the timing. It disappears in debug mode.

Sequential testing is structurally blind to them. Unit tests, functional tests, and manual QA all exercise one operation at a time. Race conditions only exist when two operations overlap. You can have 100% test coverage and 0% race condition coverage at the same time. The tests aren’t wrong. They’re measuring the wrong dimension.

Static analysis mostly can’t reason about them. SAST tools catch SQL injection and XSS because those have detectable syntactic patterns. Race conditions require reasoning about timing across concurrent executions, which static analysis can’t do in the general case. The code looks correct in isolation. It just isn’t correct when two copies run at once.

Code review catches operations, not interactions. A race condition lives in the gap between two correct operations. Each operation, reviewed on its own, passes. The bug only exists in the overlap. A reviewer who approves both operations individually has done their job correctly and still shipped the vulnerability.

How to Fix a Race Condition

There’s no single universal fix. The right approach depends on where the race window lives and what your system looks like. Here are a few fixes ordered by reliability:

1. Atomic database operations: Collapse the check and the act into a single statement. UPDATE inventory SET qty = qty - 1 WHERE id = 1 AND qty > 0 is atomic; the database guarantees no concurrent transaction slips between the condition and the write. Check the affected row count. Zero rows means the condition failed, and you return “out of stock” instead of overselling. No application-level coordination required. For web application race conditions, this is the highest-reliability fix available.

2. SELECT FOR UPDATE (pessimistic locking). When the operation is too complex for a single atomic statement, lock the row at read time. No concurrent transaction can modify that row until the lock releases. Reliable for single-database architectures, at the cost of latency under high contention. For financial and inventory operations, also raise the isolation level to REPEATABLE READ or SERIALIZABLE. Several major databases default to READ COMMITTED, which permits non-repeatable reads, the root cause of most web application race conditions.

3. Optimistic locking with a version column. Add a version integer to the table. Read it with the data, then include it in the update: UPDATE ... WHERE id = 1 AND version = 5. Zero rows affected means another transaction got there first, so you retry or return a conflict. No lock held, conflict detected at write time. Best for low-contention workloads where retries are acceptable.

4. Idempotency keys. For operations a client might retry (payments, transfers, webhook delivery), require a unique key per logical request. Store it on first processing and return the cached result for duplicates. This prevents duplicate processing regardless of race timing or retry behavior.

5. Queue-based serialization. For high-throughput scenarios, route updates through a message queue with a single consumer per logical item. Serial processing eliminates the race window entirely. Pair it with idempotency keys at the consumer, since most queues deliver at-least-once.

6. Database constraints as the last line of defense. Unique constraints, check constraints like qty >= 0, and foreign keys don't prevent race conditions. What they do is turn silent data corruption into a hard database error you'll see in your logs. Add them anyway, always. They’re the crash net, not the tightrope.

5 Questions to Find Race Conditions in Your Own Codebase Right Now

You don’t need a formal audit to start. These five questions will surface most of the exposure:

  1. Is there a check-then-act pattern? Any SELECT followed by a conditional UPDATE in separate statements is a candidate. Mentally execute it twice simultaneously with the same input. If the second execution can see the state before the first one commits, you have a race window.
  2. What happens if this endpoint receives 50 identical requests in 100ms? The cheap version: open two browser tabs and hit the same “redeem” or “purchase” button at the same time. For financial or inventory endpoints, run a proper concurrent load test before shipping, not after a user reports the bug.
  3. Does any counter, balance, quantity, or boolean flag get read before it’s written? These are the highest-value targets. If the read and the write aren’t in the same atomic operation, they’re vulnerable under concurrent load.
  4. Are uniqueness constraints enforced at the database layer? Application-level checks like if email not in database: insert are always raceable. A unique constraint at the database layer is not. If your uniqueness guarantee lives only in application code, move it down a layer.
  5. Do any privileged processes check a file path before using it? Any exists() then open(), or access() then fopen(), in a process with elevated privileges is a potential TOCTOU (Time of Check to Time of Use). Drop the check and handle the exception from the operation itself.

TestFiesta Makes Test Management Easy So You Ship Quality Software

Everything above points at one structural fact: race conditions survive standard test suites not because the tests are bad, but because sequential test execution is the wrong instrument for a concurrency problem. A suite that runs one operation at a time cannot, by definition, exercise the overlap where these bugs live.

TestFiesta closes that gap at the test management layer. Structured concurrent test execution, test case tracking across parallel runs, and coverage visibility that shows your team exactly which critical paths have never been tested under concurrent load. Because you can’t fix what you haven’t measured, and you can’t measure race condition exposure with a suite built to run one thing at a time.

If your tests aren’t simulating concurrency, they aren’t testing your system’s actual behavior.

Don’t wait for a race condition to show up in your logs. Identify, test, and resolve concurrent vulnerabilities today.

Start Your Free Trial

FAQs

Is a race condition always a security vulnerability?

Not always. In business logic (inventory, balances, coupons), it’s a data integrity bug with financial consequences. In security-sensitive paths like permission checks or privileged file operations, it becomes exploitable, so what the race window touches determines which one you have.

Do race conditions only happen in multi-threaded applications?

No. The most common race conditions in web applications happen between separate HTTP requests hitting the same endpoint at once, with no threads involved. Even single-threaded Node.js creates race windows through async/await, and serverless handlers running in parallel are especially prone.

Can automated tools reliably detect race conditions?

Only partially. ThreadSanitizer and Go's -race flag catch data races reliably, but not semantic race conditions where the logic is wrong despite synchronized memory access. The most reliable detection is deliberate concurrent testing: fire dozens of simultaneous requests at sensitive endpoints and watch for constraint violations in production logs.

Tool

Pricing

TestFiesta

Free user accounts available; $10 per active user per month for teams

TestRail

Professional: $40 per seat per month

Enterprise: $76 per seat per month (billed annually)

Xray

Free trial; Standard: $10 per month for the first 10 users (price increases after 10 users)

Advanced: $12 per month for the first 10 users (price increases after 10 users)

Zephyr

Free trial; Standard: ~$10 per month for first 10 users (price increases after 10 users)

Advanced: ~$15 per month for the first 10 users (price increases after 10 users)

qTest

14‑day free trial; pricing requires demo & quote (no transparent pricing)

Qase

Free: $0/user/month (up to 3 users)

Startup: $24/user/month

Business: $30/user/month

Enterprise: custom pricing

TestMo

Team: $99/month for 10 users

Business: $329/month for 25 users

Enterprise: $549/month for 25 users

BrowserStack Test Management

Free plan available

Team: $149/month for 5 users

Team Pro: $249/month for 5 users

Team Ultimate: Contact sales

TestFLO

Annual subscription (specific amounts per user band), e.g., Up to 50 users: $1,186/yr; Up to 100 users: $2,767/yr; etc.

QA Touch

Free: $0 (very limited)

Startup: $5/user/month

Professional: $7/user/month

TestMonitor

Starter: $13/user/month

Professional: $20/user/month

Custom: custom pricing

Azure Test Plans

Pricing tied to Azure DevOps services (no specific rate given)

QMetry

14‑day free trial; custom quote pricing

PractiTest

Team: $54/user/month (minimum 5 users)

Corporate: custom pricing

Black Box Testing

White Box Testing

Coding Knowledge

No code knowledge needed

Requires understanding of code and internal structure

Focus

QA testers, end users, domain experts

Developers, technical testers

Performed By

High-level and strategic, outlining approach and objectives.

Detailed and specific, providing step-by-step instructions for execution.

Coverage

Functional coverage based on requirements

Code coverage

Defects type found

Functional issues, usability problems, interface defects

Logic errors, code inefficiencies, security vulnerabilities

Limitations

Cannot test internal logic or code paths

Time-consuming, requires technical expertise

Aspect

Test Plan

Test Case

Purpose

Defines the overall testing strategy, scope, and approach for a project or release.

Validates that a specific feature or functionality works as expected.

Scope

Covers the entire testing effort, including what will be tested, resources, timelines, and risks.

Focuses on a single scenario or functionality in the broader scope.

Level of Detail

High-level and strategic, outlining approach and objectives.

Detailed and specific, providing step-by-step instructions for execution.

Audience

Project managers, stakeholders, QA leads, and development teams.

QA testers and engineers.

When It's Created

Early in the project, before testing begins.

After the test plan is defined and the requirements are clear.

Content

Scope, objectives, strategy, resources, schedule, environment details, and risk management.

Test case ID, title, preconditions, test steps, expected results, and test data.

Frequency of Updates

Updated periodically as project scope or strategy changes.

Updated frequently as features change or bugs are fixed.

Outcome

Provides direction and clarifies what to test and how to approach it.

Produces pass or fail results that indicate whether specific functionality works correctly.

Tool

Key Highlights

Automation Support

Team Size

Pricing

Ideal For

TestFiesta

Flexible workflows, tags, custom fields, and AI copilot

Yes (integrations + API)

Small → Large

Free solo; $10/active user/mo

Flexible QA teams, budget‑friendly

TestRail

Structured test plans, strong analytics

Yes (wide integrations)

Mid → Large

~$40–$74/user/mo)

Medium/large QA teams

Xray

Jira‑native, manual/
automated/
BDD

Yes (CI/CD + Jira)

Small → Large

Starts ~$10/mo for 10 Jira users

Jira‑centric QA teams

Zephyr

Jira test execution & tracking

Yes

Small → Large

~$10/user/mo (Squad)

Agile Jira teams

qTest

Enterprise analytics, traceability

Yes (40+ integrations)

Mid → Large

Custom pricing

Large/distributed QA

Qase

Clean UI, automation integrations

Yes

Small → Mid

Free up to 3 users; ~$24/user/mo

Small–mid QA teams

TestMo

Unified manual + automated tests

Yes

Small → Mid

~$99/mo for 10 users

Agile cross‑functional QA

BrowserStack Test Management

AI test generation + reporting

Yes

Small → Enterprise

Free tier; starts ~$149/mo/5 users

Teams with automation + real device testing

TestFLO

Jira add‑on test planning

Yes (via Jira)

Mid → Large

Annual subscription starts at $1,100

Jira & enterprise teams

QA Touch

Built‑in bug tracking

Yes

Small → Mid

~$5–$7/user/mo

Budget-conscious teams

TestMonitor

Simple test/run management

Yes

Small → Mid

~$13–$20/user/mo

Basic QA teams

Azure Test Plans

Manual & exploratory testing

Yes (Azure DevOps)

Mid → Large

Depends on the Azure DevOps plan

Microsoft ecosystem teams

QMetry

Advanced traceability & compliance

Yes

Mid → Large

Not transparent (quote)

Large regulated QA

PractiTest

End‑to‑end traceability + dashboards

Yes

Mid → Large

~$54+/user/mo

Visibility & control focused QA

Related Articles

Introduction

Accessibility testing checks whether people with disabilities can use your website or app without barriers. This includes users who are blind or low-vision, deaf or hard of hearing, keyboard-only navigators, and those with cognitive or learning disabilities.

While most teams value accessibility, many struggle to define what to test, select the right tools, or maintain checks in their release workflow. This guide covers the essential standards of accessibility testing, a three-layer testing methodology, key web and mobile differences, top tools, and answers to common team questions.

What Is Accessibility Testing

Accessibility testing is the process of checking that a digital product works for everyone, including people who rely on assistive technology or use their devices in different ways. It asks simple questions, such as:

  • Can someone reach every button using only a keyboard? 
  • Does a screen reader announce what an image shows? 
  • Is the text readable against its background? 
  • Can someone who can’t hear follow a video?

Accessibility testing is continuous, and it works best when it runs alongside the rest of your QA process.

Why Accessibility Testing Matters

There are three good reasons to take accessibility testing seriously:

  • More people can use your product: Disabilities are common, and many more people deal with temporary or situational limits, like a broken arm, bright sunlight on a screen, or a noisy room.
  • Legal risk goes down: Many countries have laws that require digital products to be accessible, and most of them point to the same set of guidelines.
  • Quality goes up: Clear labels, logical structure, and good contrast make a product easier for everyone, not just users with disabilities.

What Are Web Content Accessibility Guidelines (WCAG)?

The Web Content Accessibility Guidelines, or WCAG, are the global standard for digital accessibility. They’re published by the World Wide Web Consortium (W3C), and most accessibility laws and audits measure products against them. 

WCAG is built on four principles, often shortened to POUR:

  • Perceivable: Users can see or hear the content. Images have text alternatives, videos have captions, and text has enough contrast.
  • Operable: Users can use every control, whether they navigate with a mouse, keyboard, or voice.
  • Understandable: Content and behavior are clear and predictable. Forms explain errors, and navigation stays consistent.
  • Robust: Content works with a wide range of browsers and assistive technologies.

Under these principles sit testable rules called success criteria. Each one has a conformance level:

Level A (Minimum Compliance): Establishes the essential baseline for web accessibility. It addresses fundamental barriers that would otherwise prevent users with disabilities from accessing or interacting with content entirely (e.g., providing alt text for images, basic keyboard navigation, and avoiding content that causes seizures).

Level AA (Standard Compliance): Serves as the widely accepted global benchmark for web accessibility and legal compliance. It removes significant barriers for a broader population of users with diverse visual, hearing, cognitive, and motor impairments (e.g., ensuring adequate text contrast, visible focus states, consistent navigation, and clear error handling).

Level AAA (Enhanced Compliance): Represents the highest and most stringent level of accessibility. It incorporates specialized criteria designed to optimize usability for users with specific or severe disabilities (e.g., providing sign language interpretation for audio, extended color contrast ratios, and simplified reading levels). Note that full Level AAA compliance across an entire site is generally not required or feasible for all content types.

The Accessibility Testing Tools Worth Knowing

Instead of using dozens of accessibility tools and cluttering your QA process, focus on a core set that covers each layer of testing. Most QA teams rely on the following essentials.

axe DevTools: Scans pages for WCAG issues and explains how to fix each one.

WAVE: Places icons directly on the page to show errors, alerts, and structure. 

Google Lighthouse: Gives each page an accessibility score and a list of issues. 

Pa11y: Command-line tool that runs accessibility checks on any URL. 

Accessibility Scanner (Android): A free app that scans any screen on an Android device and suggests fixes for issues.

Xcode Accessibility Inspector (iOS): A tool that comes with Xcode and lets you inspect the accessibility details of each element in an iOS app.

Learn more about accessibility testing tools.

TestFiesta Keeps Your Accessibility Testing From Falling Through the Cracks

Accessibility testing often fails due to fragmented workflows: automated scans stay in CI/CD, manual logs remain in personal notes, and screen reader findings get lost in chat. TestFiesta unifies all three testing layers into a single platform.

  • Reusable Test Cases: Convert keyboard, zoom, and screen reader checks into standardized test cases with reusable steps for repeated actions.
  • Test Execution and Tracking: Schedule accessibility passes per release to monitor pass/fail status and test coverage in real time.
  • Defect Management: File accessibility defects directly from test runs and sync them to Jira or GitHub with full contextual details.
  • AI Test Case Generation: Automatically generate accessibility test cases from requirements, user stories, or UI screenshots.
  • CI/CD Integration: Import automated test results via the TacoTruck CLI alongside manual testing reports.

Accessibility shouldn’t be an afterthought or a chaotic scramble before release. Take control of your accessibility testing with TestFiesta.

Start a free trial today

FAQs

How is accessibility testing different from usability testing?

Usability testing checks whether a product is easy and pleasant to use for a general group of users. Accessibility testing checks whether people with disabilities can use it at all, often against WCAG criteria. 

Is accessibility testing different from user acceptance testing?

Yes, User acceptance testing (UAT) confirms that a product meets business requirements before release, usually with end users or stakeholders. Accessibility testing focuses on whether people with disabilities can use the product. 

Which WCAG level should my team be targeting, A, AA, or AAA?

Most teams should target WCAG 2.2 Level AA. It’s the level most laws and contracts ask for, and it covers the barriers users hit most often. 

Can I pass an accessibility audit using only automated tools?

No, automated tools only catch part of the WCAG criteria, and they can’t judge things like whether alt text makes sense or whether a screen reader user can complete a task. Auditors test manually and with assistive technology, so you should too.

Testing guide
Best practices

Introduction

Large language models are now a part of core enterprise infrastructure, which is why it’s critical to ensure they are safe, accurate, and high-performing at scale. 

Traditional software testing methods fall short in this landscape. As a result, QA teams are required to adopt entirely new frameworks, metrics, and red-teaming strategies to build truly production-ready AI applications.

What Is LLM Testing

LLM testing is the process of evaluating a large language model or LLM-powered application against defined expectations for quality, safety, performance, and security. In practice, most QA teams are not testing a foundation model from scratch. They are testing what sits on top of it, such as prompts, retrieval pipelines, tool calls, guardrails, and business logic that turn a general model into a product. That distinction shapes the work. 

A chatbot that answers billing questions needs to pull the right account policy, stay on topic, refuse requests it should not fulfill, and respond fast enough that customers don’t give up on it. In this case, each function a chatbot performs is a testable behavior.

LLM testing usually combines three kinds of checks:

  1. Deterministic assertions for anything with a clear right answer, such as valid JSON, required fields, or a banned phrase never appearing.
  2. Model-graded evaluation, where a second model scores outputs against a rubric. This is often called LLM-as-a-judge.
  3. Human review for judgment calls that automated scoring still gets wrong.

The result of LLM testing is a measured picture of how often the system behaves as intended, and whether it improves or degrades with each change.

The 4 Main Types of LLM Testing

Most LLM test plans fall into four categories, each targeting a different risk. A production-ready application needs coverage in all four. 

1. Functional Testing

Functional testing checks whether the application does its job, like answering correctly, following instructions, and producing output in the expected format.

Typical test cases cover:

  • Task accuracy: Does the summary capture the key points? Does the extraction pull the right fields?
  • Instruction following: Does the model respect length limits, tone requirements, and output schemas?
  • Retrieval quality: In RAG systems, did the pipeline fetch the right documents, and did the answer actually use them?
  • Multi-turn behavior: Does the assistant keep track of context from earlier in the conversation?
  • Tool and agent behavior: Did the agent call the right tool, with valid arguments, in a sensible order?

Build functional tests from real user scenarios rather than generic prompts. A test set drawn from actual support tickets or product requirements catches far more than one built from templated questions. Every bug a user reports should also become a permanent regression case.

2. Performance and Reliability Testing

Performance testing measures how fast and how consistently the application responds under realistic load. LLM latency behaves differently from typical API latency because responses stream token by token and output length varies with every request.

Key areas to cover:

  • Latency: Time to first token, total response time, and tokens per second.
  • Load behavior: How latency and error rates change as concurrent users grow, and where provider rate limits kick in.
  • Consistency: Whether the same prompt produces stable quality across repeated runs.
  • Failure handling: Timeouts, provider outages, truncated responses, malformed output, and whether retries and fallbacks work.
  • Cost: Token usage per request, since a prompt change that doubles output length also doubles spend.

Reliability is where non-determinism hurts most. Run each critical test case several times and track the spread, not just the average. A feature that answers correctly 9 times out of 10 still fails visibly for a share of real users.

3. Safety and Content Testing

Safety testing checks whether the application produces harmful, biased, or off-policy output, including when users are not trying to break it.

Coverage usually includes:

  • Toxicity and harmful content: Hate speech, harassment, and violent or sexual content.
  • Bias: Whether answers or tone shift based on names, genders, or other demographic cues in the prompt.
  • Policy compliance: Staying within brand guidelines and avoiding medical, legal, or financial advice where the product shouldn't give it.
  • Refusal calibration: Refusing what it should without over-refusing legitimate requests.
  • Privacy: Not repeating personal data from context or logs in its responses.

Over-refusal deserves its own test cases. A model that declines half of valid customer questions passes every safety check and still fails the product. Track refusal rate and harmful output rate together.

4. LLM Penetration Testing

LLM penetration testing attacks the application the way a malicious user would. Safety testing asks whether the model says something harmful, whereas penetration testing asks whether an attacker can make the system do something it was never meant to do.

Common attack categories to test:

  • Direct prompt injection: Inputs that try to override system instructions, such as “ignore previous instructions and…”
  • Indirect prompt injection: Malicious instructions hidden in content the model reads, like a web page, uploaded file, or document in a RAG index.
  • System prompt leakage: Attempts to extract hidden instructions, internal rules, or credentials placed in prompts.
  • Sensitive data disclosure: Getting the model to reveal other users’ data or internal documents from its context.
  • Excessive agency: Tricking an agent into calling tools, sending messages, or changing data beyond what the user is authorized to do.
  • Insecure output handling: Model output passed unchecked into a browser, database query, or shell.
  • Resource abuse: Prompts designed to max out tokens, loop agents, or run up costs.

When a model can take actions, a successful injection can send emails, modify records, or expose data. Pair automated attack generation with manual testing, and retest after every prompt or tool change.

LLM Testing Frameworks and Tools Worth Knowing

No single tool covers every type of LLM testing. Most teams pair an evaluation framework for functional and RAG metrics with a dedicated tool for security scanning. 

1. DeepEval

DeepEval is a Python LLM testing framework (Apache 2.0) integrated with Pytest. It provides metrics for RAG quality, agent behavior, conversation context, toxicity, custom evaluation, and synthetic dataset generation for CI/CD integration.

Best for: Teams that want LLM tests to live alongside their existing Python test suite.

2. Promptfoo

Promptfoo is a CLI tool and library (MIT) for declarative evaluation and red-teaming. It compares prompts and models across providers and generates adversarial tests for prompt injection, jailbreaks, and data exfiltration.

Best for: Prompt and model comparisons, and automated LLM penetration testing.

3. Giskard

Giskard is an open-source Python library (Apache 2.0) that automatically detects hallucinations, prompt injection, and bias. Its RAGET module generates test cases to evaluate retriever and generator performance separately.

Best for: Surfacing vulnerabilities you don’t typically test for, especially in RAG applications.

4. TruLens

TruLens (MIT) pairs OpenTelemetry tracing with model evaluations to measure latency, token usage, cost, and the RAG triad (context relevance, groundedness, and answer relevance).

Best for: Pinpointing where in a multi-step pipeline quality breaks down, and comparing app versions on quality, latency, and cost.

5. Ragas

Ragas (Apache 2.0) evaluates retrieval-augmented generation systems through faithfulness, answer relevancy, context precision, and recall metrics, alongside synthetic test dataset generation.

Best for: RAG evaluation and teams that need to build a test set from scratch.

Key LLM Testing Metrics

LLM testing differs from usual software testing metrics. Most quality metrics in LLM testing score from 0 to 1, and your team sets the threshold that counts as a pass.

1. Faithfulness

Faithfulness measures whether the claims in an answer are supported by the context the model was given. It is usually calculated by breaking the answer into individual statements and checking each one against the retrieved documents. A low score means the model is adding information its sources don't contain.

2. Answer Relevance

Answer relevance measures how directly a response addresses the question asked. An answer can be accurate and faithful but still miss the point, filling space with related facts instead of answering. This metric catches that.

3. Hallucination Rate

Hallucination rate is the share of outputs that contain fabricated or unsupported information. Faithfulness scores a single response, while hallucination rate is tracked across a full test set, so you can see whether a model or prompt change made the problem better or worse.

4. Toxicity Score

A toxicity score estimates how likely an output is to be offensive, abusive, or harmful, usually using a classifier or judge model. Track the share of responses above your threshold rather than the average score. One toxic reply matters more than a low mean.

5. Groundedness

Groundedness measures how well each statement in a response can be traced back to a provided source. It overlaps heavily with faithfulness, and frameworks define the two terms differently. Check how your tool calculates each before comparing numbers across tools.

6. Latency / TTFT (Time to First Token)

Time to first token is the delay between sending a request and receiving the first token of the response. It drives how fast a streaming interface feels. Track it alongside total response time and tokens per second, and report percentiles such as p95 instead of averages, because slow outliers are what users notice.

7. Pass@k

Pass@k is the probability that at least one of k generated attempts passes a test. It comes from code generation, where a task counts as solved if any of k samples passes its unit tests. For user-facing features, the stricter question is often whether all k attempts pass, since users see one response, not the best of several.

TestFiesta Was Built for the Complexity LLM Testing Brings

LLM testing creates a lot of moving parts, including functional, safety, and security test cases, results from automated eval runs, scores that shift with every prompt or model change, and defects that need a clear trail back to the input that caused them. TestFiesta gives QA teams one place to manage all of it.

  • AI Copilot for Test Design: Upload requirements, user stories, or screenshots, and the Copilot generates structured test cases from that context. Nothing enters your suite until you review and approve it, which matters when you are writing hundreds of prompt scenarios and adversarial cases.
  • Flexible Organization: Reusable templates, universal tagging, and shared steps let you group cases by test type, risk, or feature, so a safety suite or red-team suite can run on its own.
  • Reusable Configurations: Run the same test cases across environments, or treat model and prompt versions as configurations to track results per version.
  • Automated Results in One Place: The TacoTruck CLI and full API access push results from CI/CD runs into TestFiesta, so eval framework output sits next to manual and exploratory testing.
  • Built-In Bug Tracking: Log defects in the same platform where you manage test cases, or turn a failed test into a Jira or GitHub issue with full context attached.

Set up your first LLM test suite in TestFiesta and get a QA workflow that fits your team.

Start your free trial today

FAQs

Can LLM testing be automated?

Yes, LLM testing can be largely automated. Deterministic checks, model-graded metrics, and adversarial attack generation can all run in a CI/CD pipeline on every prompt, model, or code change. That said, automation can’t fully replace human review of judgment-heavy outputs, or periodic checks that judge whether models are scoring accurately. 

What is LLM red teaming?

LLM red teaming is adversarial testing where testers deliberately try to make a model or application misbehave through jailbreaks, prompt injection, data extraction, or harmful content requests. It overlaps with LLM penetration testing but is usually broader, covering safety and misuse as well as security flaws. 

How do I test an LLM when I don’t have a ground truth dataset?

You can still test an LLM if you don’t have a ground truth dataset. Start with reference-free metrics such as faithfulness, answer relevance, and groundedness, which score outputs against the question and retrieved context instead of an expected answer. You can also generate a synthetic test set from your own documents and have domain experts review a sample. Over time, real user queries from production, labeled as you go, become your ground truth dataset.

Testing guide
Best practices

Introduction

How do you verify whether the system output is actually correct when nobody knows what the exact expected answer should be?

In software engineering, this challenge is known as the oracle problem; a bottleneck where traditional unit tests and pass/fail assertions hit a wall. When checking single outputs against fixed expected values is impossible or too expensive, metamorphic testing offers a powerful alternative: instead of checking whether a single output is correct on its own, you check how outputs relate across logical transformations of the input.

What Is Metamorphic Testing

Metamorphic testing is a way to test software without knowing the exact correct output for any single input. Instead of checking one result against an expected value, you run the program more than once with related inputs. Then you check that the outputs relate to each other the way they logically should.

Here is a simple illustration:

Say you are testing a payroll function. You may not know the exact total for a 400-person company with overtime, bonuses, and tax rules. But you do know one thing for sure: If every employee’s hourly rate doubles and nothing else changes, gross payroll should double too. Run both versions, compare, and if the second total isn’t twice the first, you can determine that the output is wrong. As a result, you found a problem without ever knowing the “right” answer.

Metamorphic testing has two main strengths:

  • It eases the oracle problem: You verify relationships between outputs, not the correctness of each output on its own.
  • It generates new tests: Every existing test case can spawn follow-up cases, so coverage grows from tests you already have.

What Is a Metamorphic Relation

A metamorphic relation (MR) is the rule that ties metamorphic test runs together. It states how a specific change to the input should affect the output. If the outputs break the rule, the program has a fault somewhere.

Every MR has two parts. One is the input transformation, which is how you change the original input to create a new one. The other is the output relation, which is what must hold between the original output and the new output.

Imagine you are testing a function that sorts a list of numbers. For a short list, you can check the result by looking at it. But for a list of 100,000 numbers, you need a function. When testing complex functions like sorting a massive list of numbers where you cannot manually verify every expected output, you can still rely on fundamental rules or properties that must always remain true. Each of these inherent properties serves as a Metamorphic Relation (MR) that can be used to catch software bugs without needing to know the exact final answer in advance.

In this specific case, the MRs would be:

MR 1: Shuffling the input should not change the output. If list B contains exactly the same numbers as list A in a different order, then sort(B) must equal sort(A). If they differ, the function is mishandling order somewhere.

MR 2: Adding a larger number should only extend the output. If you append a value bigger than everything in list A, the sorted result should be the original sorted list with that value at the end. Anything else means the new element disturbed items it never should have touched.

Neither rule tells you what the sorted list should look like. But both can still catch real bugs. Mathematical functions work the same way. You may not know sin(x) to fifteen decimal places, but you know sin(x) and sin(π − x) must be equal.

How Metamorphic Testing Works: A Step-by-Step Guide

Metamorphic testing follows the five steps, which remain the same regardless of testing a sort function or a self-driving model. 

To make it simple, we’ll use the e-commerce search example.

Step 1: Identify metamorphic relations: Study what the system is supposed to do and write down properties that must hold across related inputs, like “adding a filter never adds results.” This step takes the most thought and depends on domain knowledge, so pull in the people who understand the system best.

Step 2: Create source test cases. Choose the original inputs the process starts from, such as a set of real search queries. These can be existing test cases, randomly generated inputs, or anonymized production data.

Step 3: Derive follow-up test cases. Apply each MR’s input transformation to every source case, such as by adding a size filter to each query. One source case can feed several MRs, so a small seed set grows quickly.

Step 4: Execute both sets. Run the program on the source inputs and the follow-up inputs, then record every output. Keep each follow-up paired with its source so you always know which results to compare.

Step 5: Verify the relation. Check whether each pair of outputs satisfies the output relation, such as “filtered results are a subset of unfiltered results.” A violation means at least one output is wrong, which signals a bug in the program or a flaw in the MR itself.

A Pro Tip: A passing metamorphic test does not prove the program is correct. A function could break in a way that still satisfies a given relation. That is why teams usually define several MRs per feature, each checking the behavior from a different angle.

Where Metamorphic Testing Matters

Metamorphic testing matters wherever expected outputs are unknown, expensive to compute, or open to interpretation. These five areas are where it shows up most include:

Scientific and mathematical software

  • Simulations and numerical programs that often have no known exact answer to compare against.
  • Mathematical properties like symmetry, scaling, and conservation laws.
  • Published studies have applied it to bioinformatics programs, epidemiological models, and software at NASA.

Machine learning and AI models

  • Models that produce outputs that no one can fully specify in advance, which is the oracle problem in its purest form.
  • Machine learning models where output should remain consistent or stable when minor, non-essential changes are made to the input. In this context, if you slightly darken a photo, an image classifier should still recognize the same object, because a minor change in lighting doesn’t change the underlying subject of the image.
  • Large language models where rephrasing a factual question should not change the factual answer.

Search engines and ranking systems

  • Systems where there is no objective, agreed-upon “correct” result set for most queries.
  • Subset and consistency relations, like the filtering examples.

Compilers and code translation tools

  • Tools where verifying the compiled output for every possible program is out of reach.
  • Translation tools where two programs that behave the same on a given input should still behave the same after compilation.

Web APIs

  • API responses that often depend on live, constantly changing data, so exact expected values are hard to pin down.

How TestFiesta Helps With Test Case Management for Metamorphic Testing

Metamorphic testing multiplies your test cases fast. Five MRs applied to 40 source cases give you 200 follow-up cases, and every one of them needs to stay linked to its source, its relation, and its result. Without structure, that turns into a cluttered dashboard or spreadsheet.

TestFiesta keeps everything around them organized and traceable.

  • Customizable Tags: TestFiesta offers unlimited customizable tags for grouping, filtering, and sorting test cases. Tag each case with its MR and its role as a source or follow-up case. When an MR starts failing, you can filter straight to every related case.
  • Custom Fields: Custom fields capture data at the test case, run, or result level. Use them to record the input transformation, the expected output relation, and the ID of the source case each follow-up came from.
  • Reusable Templates: Build one MR template with fields for the transformation and the relation. Every new MR then gets documented the same way, no matter who on the team writes it.
  • Shared Steps: Source and follow-up cases usually share setup, like loading a dataset or authenticating against an API. Write those steps once and reuse them everywhere, so a change in setup updates every case at once.
  • Configurations: Configurations let you reuse one test case across many setups. Run the same MR against different environments, browsers, or model versions without cloning cases.
  • Jira and GitHub Integrations: When a relation is violated, turn the failed test into a Jira or GitHub issue with full context, or log bugs natively in TestFiesta. Developers see both inputs and both outputs, which is exactly what they need to debug a metamorphic failure.
  • An API for Automated Results: Metamorphic tests are usually automated, and TestFiesta’s Automation API supports custom automations, so results from your test runs can flow into the platform instead of being copied by hand.

Don’t let the complexity of managing metamorphic relations and automated test suites slow down your release cycles.

Start your free trial today

FAQs

Can metamorphic testing be automated?

Yes, metamorphic testing can be automated, and it is well-suited to automation. Once your MRs are defined, generating follow-up inputs, running both sets, and checking the relations can all be scripted. 

What tools do I need for metamorphic testing?

You don’t need a dedicated tool for metamorphic testing. Most teams write MRs as ordinary tests in the framework they already use, such as pytest or JUnit, and property-based testing libraries like Hypothesis can help generate source inputs. A test case management platform keeps your MRs, source cases, follow-ups, and results connected as the suite grows.

Do I need to know the correct output to write metamorphic tests?

No, you don’t need to know the correct output to write metamorphic tests. That’s the entire point of metamorphic testing. You only need to know how outputs should relate to each other when the input changes in a specific way. Writing good relations still takes domain knowledge, but you never need the exact expected value for any single input.

Testing guide

Ready for a Platform that Works

The Way You Do?

Stop fighting your tools. Start shipping with confidence. TestFiesta adapts to your workflow, not the other way around.

Welcome to the fiesta!