Lesson 34 of 35AI Red TeamingAI Red Teaming

What is AI Red Teaming?AI Security Testing & Model Assurance explained

Tests AI models and applications for prompt injection, jailbreaks, data leakage, unsafe behavior, and model supply-chain risks before and after deployment.

By Reviewed

Verified Sources: csrc.nist.gov, atlas.mitre.org, owasp.org

The lesson

How defenders break an AI system before attackers do

AI behavior changes with prompts, models, retrieved content, tools, and updates, so a single vulnerability scan cannot establish safety. This lesson explains adversarial testing and model supply-chain assurance.

If it helps, think of it as… crash tests and parts inspection

A car maker checks both the parts arriving at the factory and the finished car in a crash test. AI assurance works the same way: scan model artifacts and dependencies before they enter the pipeline, then challenge the assembled application with hostile prompts and realistic attack paths. A clean part does not prove a safe car, and one successful crash test does not cover every future model update.

Inspect artifacts

Models, code, lineage, integrity, dependencies

Define threats

Prompt injection, leakage, theft, abuse

Run attacks

Automated probes plus expert red teaming

Score failures

Reproducible evidence, not a safety claim

Fix and retest

Regression tests follow every material change

…then the loop starts again — this runs continuously, not once.

Assurance repeats across the lifecycle: inspect components, test assembled behavior, fix failures, and rerun whenever the system changes.

Test the parts and the behavior

AI assurance has two linked jobs. First, inspect the artifacts and dependencies entering the system. Second, attack the behavior of the assembled model or application through the interfaces a real user or adversary can reach.

Model files deserve special handling because some serialization formats can execute code when loaded. Scanning before use reduces that risk, while hashes, signatures, provenance, and access controls help establish whether an artifact is the one the team intended to deploy.

Build a threat model

A useful test starts with the system, its assets, and its trust boundaries. Is the target a public chatbot, an internal assistant with sensitive retrieval, or an agent that can take actions? The same prompt can have radically different consequences in each case.

Threat libraries provide a starting point. Prompt injection, jailbreaks, data disclosure, model theft, poisoned content, and unsafe tool use are distinct problems. Test plans should state which are in scope and which are not.

Automate repeatable attacks

Automated frameworks combine targets, attack prompts, transformations, and scoring. They are valuable for breadth and regression testing: a team can rerun the same tests after changing a model, prompt, retrieval source, or tool.

Scoring needs review. An automated detector can miss context or mistake a harmless output for a failure. High-impact findings should preserve the input, output, configuration, model version, and reasoning needed for a human to reproduce the result.

Add human red teaming

Expert testers explore multi-step paths and adapt when the system resists an attack. They can combine ordinary application vulnerabilities with AI-specific behavior—for example, using a weak authorization boundary after manipulating an agent's instructions.

Human testing is deeper but less repeatable, which is why mature programs turn confirmed findings into automated regression cases.

Interpret the result correctly

A passed test means a defined attack did not succeed under defined conditions. It does not prove that the model is safe. Probabilistic behavior, new attack techniques, and frequent system changes make assurance a repeating lifecycle rather than a one-time gate.

Terms you just met

Each links to its plain-language definition in the glossary.

The field guide

Evaluating this category

A second pass for buyers: market context, distinctions that matter, and what to weigh when tools in this category start looking alike.

AI systems need security testing at two levels. The components entering the system—model files, code, dependencies, datasets, and configuration—need supply-chain assurance. The assembled model or application also needs adversarial testing against the behavior an attacker can reach.

AI security testing and model assurance combines those concerns. It includes model artifact scanning, provenance and integrity checks, automated adversarial evaluation, and expert red teaming for models, generative-AI applications, and agents.

Why conventional scanning is not enough

Traditional software testing still matters. An AI application can contain ordinary vulnerable code, exposed APIs, secrets, and dependencies. But AI adds failure modes that are expressed through model behavior and the surrounding data flow: prompt injection, jailbreaks, sensitive-data disclosure, model extraction, inference attacks, poisoned retrieved content, and unsafe tool use.

NIST's adversarial machine-learning taxonomy, MITRE ATLAS, and the OWASP GenAI Security Project provide complementary ways to describe these threats. A product may cover only a subset, so broad “AI red teaming” claims should be unpacked into specific attack types and system boundaries.

Model artifact and supply-chain checks

Serialized model files can contain executable content or unsafe operators. Scanners inspect supported formats without loading them, while broader supply-chain platforms can record provenance, hashes, lineage, dependencies, and policy decisions.

Coverage is format-specific. A scanner that understands one serialization format does not automatically validate every model framework, custom operator, dataset, or hosted model. Teams should verify exactly what is inspected and what remains a manual trust decision.

Adversarial evaluation

Automated testing sends crafted inputs to a model or application and scores the resulting behavior. Useful systems make attacks reproducible, record model and prompt versions, and support regression testing after a change. Expert red teams add creativity, system knowledge, and multi-step attack paths that a fixed probe library may miss.

The target also matters. Testing a base model through a simple chat endpoint does not test the complete application with retrieval, memory, system prompts, authorization, and tools. High-impact assessments should exercise the deployed system at the boundaries an attacker can actually reach.

What to compare

Compare supported targets, attack libraries, custom test authoring, scoring, evidence, and CI/CD integration. Ask whether tests cover models, retrieval pipelines, multimodal input, agents, and tool calls. Determine whether the product separates security failures from broader quality and safety evaluations.

Results should be treated as evidence from a defined test—not a certification that an AI system is safe. Models are probabilistic, and changes to prompts, context, providers, tools, or data can alter outcomes. Continuous evaluation is valuable because the target keeps moving.

Capability taxonomy

What buyers typically evaluate when comparing tools in this category.

AI red teaming
Runs adversarial tests against models and applications to uncover exploitable or unsafe behavior.
Prompt-injection testing
Tests whether untrusted instructions can override intended system behavior or controls.
Jailbreak & safety testing
Evaluates resistance to policy bypasses, harmful outputs, and other unsafe behaviors.
Model artifact scanning
Inspects model files and serialized artifacts for malicious code, tampering, and unsafe components.
AI supply-chain assurance
Tracks model provenance, dependencies, integrity, and risk before deployment.
Continuous AI evaluation
Repeats security tests as models, prompts, tools, data, and application behavior change.

Tools in this category

Now that you know what AI Red Teaming does, see who does it.

Search Cyber Tool Stack

Jump to any tool, vendor, category, or glossary term.