Review scope and evidence labels

This page profiles AI Red Teaming as a methodology for structured adversarial testing and abuse-case validation of AI-enabled systems.

Evidence labels used on this page:

  • Community security knowledge base — based on public MITRE ATLAS or similar adversary-knowledge material.
  • Community security guidance — based on public security community material such as OWASP.
  • Official guidance — based on public official guidance such as NIST material.
  • Author analysis — interpretation for AI security, governance, architecture and operational risk on this site.
  • Implementation evidence required — local records that must exist before a mitigation can be treated as operating.

This entry is a Research methodology profile, not a certification claim. It does not imply offensive AI red-team certification; it defines the governance and architecture questions that structured AI red-team work should answer.

Methodology snapshot

FieldValue
CategoryAI adversarial testing and control validation
Research typeMethodology
Core questionCan realistic AI abuse scenarios bypass intended controls or create unacceptable harm?
Main scopeLLM applications, RAG, agents, tools, data access, model behavior, prompts, outputs, logs and response playbooks
Useful outputsScenario library, findings, reproduction evidence, control gaps, residual risk and retest records
Companion methodsMITRE ATLAS, AI Threat Modeling, Prompt Injection, OWASP LLM Top 10, Audit Evidence
Review statusPublic guidance reviewed; local work requires explicit scope, authorization and safety boundaries

What AI Red Teaming means

AI red teaming tests whether an AI-enabled system can be manipulated, misused or made to fail in ways that matter to the organisation. Evidence: Community security knowledge base; Official guidance; Author analysis.

It should be connected to threat modeling. Random prompts are not enough; tests should map to assets, actors, abuse cases, controls and evidence. Evidence: Author analysis.

For this site, AI Red Teaming is positioned as architecture and governance validation: the point is to improve control design and risk decisions, not to claim offensive credentials. Evidence: Author analysis.

What it is not

AI Red Teaming is not:

  • a certification by itself,
  • unrestricted attacks against public systems,
  • only jailbreak prompt lists,
  • only model benchmark testing,
  • a substitute for secure design,
  • proof that a system is safe forever,
  • permission to test without written scope.

Practical AI red-team work needs authorization, test boundaries, safety rules, evidence handling and remediation follow-up. Evidence: Author analysis; Implementation evidence required.

Practical testing areas

AreaPractical questionEvidence to retain
Prompt injectionCan untrusted instructions override policy or leak data?Test cases, prompts, results, screenshots/logs
RAG abuseCan retrieved content poison context or expose restricted data?Source samples, query traces, permission tests
Tool misuseCan an agent call tools outside intended scope?Tool-call logs, approvals, negative tests
Data leakageCan sensitive context appear in answers or logs?Redaction tests, output review, log policy
Model misuseCan the system assist prohibited or harmful workflows?Safety tests, refusal analysis, policy records
MonitoringWould defenders see the abuse path?Telemetry, alerts, incident drill results
RemediationWere fixes retested and accepted?Fix records, retest evidence, residual-risk decision

Evidence: Community security guidance; Author analysis.

Relationship to existing Research entries

AI Red Teaming connects to:

  • MITRE ATLAS — provides adversary vocabulary and technique framing.
  • AI Threat Modeling — defines the scenarios that red-team tests should validate.
  • Prompt Injection — common LLM abuse path requiring specific tests.
  • RAG Security — retrieval systems need abuse tests for poisoning and leakage.
  • AI Security Architecture — red-team findings should improve architecture decisions.
  • OWASP Top 10 for LLM Applications — risk categories can become test families.
  • AI Governance — results must feed ownership, risk acceptance and remediation.
  • Audit Evidence — claims require reproducible evidence and retest records.

Governance relevance

AI red-team work is governance-relevant because it turns abstract AI risk into concrete scenarios, findings and decisions. It helps leadership see whether controls operate in practice and whether residual risk is acceptable. Evidence: Author analysis; Implementation evidence required.

Useful governance outputs include approved scope, scenario catalog, finding register, evidence pack, remediation plan, retest record and accepted-risk statement. Evidence: Author analysis.

Current limitations and follow-up research

This page is a methodology profile. Real work requires explicit permission, system access, safety constraints, evidence handling, legal scope and remediation ownership.

Follow-up Research nodes should cover:

  • AI Abuse Case Library,
  • Agent Security Testing,
  • RAG Red Teaming,
  • AI Incident Response,
  • AI evaluation governance.