Skip to content

Latest commit

 

History

History
850 lines (667 loc) · 35.9 KB

File metadata and controls

850 lines (667 loc) · 35.9 KB

LLM Top 10 2025 × PCI DSS v4.0

Mapping the OWASP Top 10 for LLM Applications 2025 to PCI DSS v4.0 — the Payment Card Industry Data Security Standard, version 4.0, published March 2022 and mandatory from March 2025.


Why PCI DSS for LLM security

PCI DSS applies to any organisation that stores, processes, or transmits cardholder data (CHD) or sensitive authentication data (SAD). If your LLM application has any of the following characteristics, PCI DSS requirements apply to the LLM systems and their data pipelines:

  • The LLM processes customer support interactions that may include payment card numbers, expiry dates, or CVV codes
  • The LLM has access to databases or document stores containing transaction records or cardholder data
  • The LLM is deployed in a payment processing or fraud detection pipeline where CHD is in scope
  • The RAG corpus for the LLM includes documentation, procedures, or records that contain cardholder data
  • Agents built on the LLM have tool access to payment systems

PCI DSS v4.0 introduced significant updates relevant to LLM applications — most importantly the requirement for customised approach, which allows organisations to meet the intent of requirements using alternative methods. For novel LLM-specific controls that do not map cleanly to traditional PCI requirements, the customised approach provides the mechanism to document them as compensating controls.

Cardholder data (CHD) in LLM scope is a critical risk. An LLM that has been trained on or has retrieval access to CHD is a PCI scope expansion that assessors will flag. The controls in this file address how to manage LLM applications in PCI scope and how to prevent inadvertent PCI scope expansion through LLM deployments.


PCI DSS v4.0 structure

PCI DSS v4.0 is organised into 12 requirements:

Requirement Title LLM relevance
1 Network security controls LLM infrastructure network segmentation
2 Secure configurations LLM and model component hardening
3 Protect stored account data CHD in training data, RAG, embeddings
4 Protect cardholder data in transit LLM API calls, RAG retrieval with CHD
5 Protect against malicious software LLM component integrity, malware prevention
6 Develop and maintain secure systems Secure development for LLM integrations
7 Restrict access to cardholder data Least privilege for LLM access to CHD
8 Identify users and authenticate access LLM service identity, authentication
9 Restrict physical access Physical controls for LLM infrastructure
10 Log and monitor all access LLM audit logging, monitoring
11 Test security systems LLM security testing programme
12 Support information security policies LLM risk management, vendor programme

PCI DSS v4.0 key new requirements for LLM

Requirement Change in v4.0 LLM application
Req 6.2 Bespoke and custom software security LLM integration code requires secure development and review
Req 6.3 Security vulnerabilities identified and addressed LLM component CVEs in vulnerability management
Req 6.4 Public-facing web applications protected LLM-powered chatbots and portals in PCI scope require WAF or equivalent
Req 6.5 Changes managed securely Model updates as system changes requiring security testing
Req 10.7 Failures of critical controls detected LLM monitoring failures treated as critical control failures
Req 12.3 Targeted risk analyses LLM-specific risk analyses documented
Req 12.6 Security awareness programme AI security awareness for staff handling CHD through LLM

Quick-reference summary

ID Name Severity Primary PCI DSS v4.0 Requirements Tier
LLM01 Prompt Injection Critical Req 6.2, Req 6.4, Req 10.6, Req 11.3 Foundational-Advanced
LLM02 Sensitive Information Disclosure High Req 3.3, Req 3.5, Req 4.2, Req 7.2 Foundational-Advanced
LLM03 Supply Chain Vulnerabilities High Req 6.3, Req 12.8, Req 2.2, Req 6.5 Foundational-Hardening
LLM04 Data and Model Poisoning Critical Req 6.5, Req 10.6, Req 11.3, Req 12.3 Hardening-Advanced
LLM05 Insecure Output Handling High Req 6.2, Req 6.4, Req 11.3, Req 10.6 Foundational-Hardening
LLM06 Excessive Agency High Req 7.2, Req 7.3, Req 10.2, Req 12.3 Foundational-Hardening
LLM07 System Prompt Leakage High Req 3.5, Req 7.2, Req 10.2, Req 6.2 Foundational-Hardening
LLM08 Vector and Embedding Weaknesses Medium Req 3.4, Req 3.5, Req 7.2, Req 11.3 Hardening-Advanced
LLM09 Misinformation Medium Req 12.3, Req 6.2, Req 10.6, Req 12.6 Foundational-Hardening
LLM10 Unbounded Consumption Medium Req 1.3, Req 6.4, Req 10.6, Req 12.3 Foundational-Hardening

Audience tags

  • CISO / compliance lead — full file, PCI DSS v4.0 alignment for LLM applications
  • QSA / assessor — requirement-to-control mapping for PCI assessment scope
  • Payment / fintech product team — CHD scope, cardholder data handling
  • Security engineer — Req 6, Req 10 technical control entries
  • Risk manager — Req 12.3 targeted risk analysis entries throughout

Detailed mappings


LLM01 — Prompt Injection

Severity: Critical

Malicious instructions manipulate LLM behaviour. In PCI scope, prompt injection against an LLM with CHD access is a critical finding — a successful attack could cause the model to exfiltrate cardholder data, generate fraudulent payment instructions, or bypass fraud detection logic.

PCI DSS assessor expectation: For LLM applications in PCI scope, Requirement 6.2 (secure software development practices) requires that injection vulnerabilities are addressed in all bespoke software including LLM integrations. Requirement 6.4 requires that public-facing applications are protected against known attacks — prompt injection is a known attack class for LLM applications and assessors will test for it.

PCI DSS v4.0 mapping

Requirement ID How it applies
Secure software development Req 6.2.4 All injection vulnerabilities addressed in LLM integration code — prompt injection as a known injection class
Public-facing application protection Req 6.4.1 LLM-powered customer-facing applications protected against prompt injection — WAF or equivalent, security testing
Audit log review Req 10.6.1 Automated log analysis covering LLM injection indicators — anomalous inputs detected and reviewed
External and internal penetration testing Req 11.3.1 Penetration testing includes prompt injection scenarios for all LLM applications in PCI scope

Mitigations for PCI DSS compliance

Foundational

  • Req 6.2.4: Implement input validation for all LLM integrations in PCI scope — prompt injection addressed as a known injection vulnerability class in secure development standards
  • Req 6.4.1: Apply web application protection to all public-facing LLM applications — WAF rules or equivalent coverage including injection detection
  • Document prompt injection as a targeted risk per Req 12.3.2 for all LLM systems in CDE scope

Hardening

  • Req 11.3.1: Include prompt injection in penetration testing scope — direct, indirect via RAG, and jailbreak vectors for all LLM applications in PCI scope
  • Req 10.6.1: Deploy automated monitoring for injection indicators — log review covers LLM-specific anomalies
  • Implement architectural separation between system prompt and user input — Req 6.2 secure design control

Advanced

  • Req 11.3.1: Conduct adversarial testing quarterly — novel injection techniques tested against LLM applications in CDE before reaching production
  • Req 12.3.2: Update targeted risk analysis when new injection techniques emerge — dynamic risk assessment
  • Customised approach: Document injection-specific controls as customised approach evidence where standard WAF rules do not cover LLM-specific vectors

Tools

Tool Type Link
Garak Open-source https://github.com/leondz/garak
LLM Guard Open-source https://github.com/protectai/llm-guard

Cross-references

  • Agentic Top 10: ASI01 Agent Goal Hijack
  • DSGAI 2026: DSGAI01 Sensitive Data Leakage, DSGAI12 Unsafe NL Data Gateways
  • Other frameworks: ISO 27001 A.8.28 · NIST AI RMF MS-2.5 · OWASP ASVS V5.1

LLM02 — Sensitive Information Disclosure

Severity: High

LLMs expose sensitive information through outputs. In PCI scope this is a primary cardholder data protection concern — an LLM that reproduces PANs, expiry dates, CVVs, or cardholder names from training data or RAG retrieval is a direct PCI DSS Requirement 3 violation.

PCI DSS critical note: CHD must never appear in LLM training data (Req 3.3, 3.4), RAG document stores (Req 3.5), or model outputs in cleartext (Req 4.2). If CHD is in LLM scope, the entire LLM infrastructure is in the CDE (Cardholder Data Environment) and full PCI DSS scope applies.

PCI DSS v4.0 mapping

Requirement ID How it applies
Protect stored account data — SAD prohibition Req 3.3.1 SAD (CVV, PIN) must never be in LLM training data or RAG — stored after authorisation is prohibited
Protect stored account data — PAN rendering Req 3.4.1 PANs in LLM outputs masked — only first six and last four digits displayed in any LLM response
Protect stored account data — encryption Req 3.5.1 Any PAN in LLM scope encrypted with strong cryptography — AES-256 or equivalent
Protect cardholder data in transit Req 4.2.1 All LLM API calls and RAG retrieval paths carrying CHD encrypted in transit — TLS 1.2 minimum
Restrict access by need to know Req 7.2.1 LLM access to CHD restricted to minimum required — retrieval access controls prevent over-broad PAN access

Mitigations for PCI DSS compliance

Foundational

  • Req 3.3.1: Audit all LLM training data and RAG corpora for SAD before any LLM in CDE scope is trained or deployed — SAD prohibition is absolute, no exceptions
  • Req 3.4.1: Implement PAN masking in all LLM outputs — automated output redaction before responses leave the LLM service boundary
  • Req 3.5.1: Encrypt all stored CHD in LLM scope — training datasets, RAG document stores, embedding stores, prompt caches containing CHD

Hardening

  • Req 4.2.1: Enforce TLS 1.2 minimum on all LLM data flows carrying CHD — API calls, RAG retrieval, observability pipelines
  • Req 7.2.1: Apply need-to-know access controls on RAG retrieval — LLM access to CHD restricted to minimum required by the specific use case
  • Scope minimisation: Design LLM applications to operate outside CDE where possible — tokenisation before data reaches the LLM removes PCI scope

Advanced

  • Apply differential privacy for any LLM trained on CHD corpora — reduces risk of PAN reproduction from training data memorisation
  • Conduct PAN detection red team — query the LLM to verify PANs cannot be reconstructed from outputs
  • Req 3.3.1: Verify SAD absence in all model artifacts — post-training audit of model outputs for SAD patterns

Tools

Tool Type Link
Microsoft Presidio Open-source https://github.com/microsoft/presidio
Nightfall AI Commercial https://nightfall.ai

Cross-references

  • Agentic Top 10: ASI03 Identity and Privilege Abuse
  • DSGAI 2026: DSGAI01 Sensitive Data Leakage, DSGAI18 Inference and Data Reconstruction
  • Other frameworks: ISO 27001 A.8.11 · NIST AI RMF GV-1.6 · SOC 2 C2.1

LLM03 — Supply Chain Vulnerabilities

Severity: High

LLM components sourced from third parties create supply chain risk. PCI DSS Requirement 12.8 (third-party service providers) applies to all vendors with access to or components that process CHD — model weight providers and dataset vendors accessing CDE data are TPSPs under PCI DSS.

PCI DSS v4.0 mapping

Requirement ID How it applies
Third-party service providers Req 12.8.1 List of all TPSPs maintained — LLM component vendors with CDE access or data in scope
TPSP agreements Req 12.8.3 Written agreements with LLM vendors — acknowledge responsibility for CHD security
TPSP PCI DSS compliance Req 12.8.4 Monitor TPSP PCI DSS compliance status — LLM component vendors with CHD access have compliant status verified
Security vulnerabilities Req 6.3.3 All LLM software components at latest security patches — ML libraries and inference runtime patched
Secure system changes Req 6.5.1 Model component updates follow change management — security testing before production

Mitigations for PCI DSS compliance

Foundational

  • Req 12.8.1: Maintain TPSP list including all LLM component vendors with CDE access or processing CHD — model providers, dataset vendors, inference platform
  • Req 12.8.3: Establish written agreements with LLM vendors acknowledging CHD security responsibility — required before any CDE-adjacent LLM deployment
  • Req 6.3.3: Maintain ML SBOM and patch all LLM components to latest security patches on schedule

Hardening

  • Req 12.8.4: Monitor LLM vendor PCI DSS compliance — obtain AOC (Attestation of Compliance) from vendors with CHD access, track compliance status annually
  • Req 6.5.1: Include model and component updates in PCI change management — security testing before any change to production LLM in CDE
  • Verify cryptographic signatures of all LLM components — unsigned components rejected before CDE deployment

Advanced

  • Operate isolated model evaluation before CDE deployment — Req 6.5 testing before change to production
  • Req 12.8.2: Document roles and responsibilities for LLM vendor CHD security — who owns TPSP oversight, what escalation path exists for vendor incidents
  • Customised approach: Document ML SBOM and component verification as a customised Req 12.8 monitoring control

Tools

Tool Type Link
CycloneDX Open-source https://cyclonedx.org
ModelScan Open-source https://github.com/protectai/modelscan

Cross-references

  • Agentic Top 10: ASI04 Agentic Supply Chain Vulnerabilities
  • DSGAI 2026: DSGAI04 Data Model and Artifact Poisoning
  • Other frameworks: ISO 27001 A.5.19/A.5.21 · NIST AI RMF MP-5.1 · NIST SP 800-218A

LLM04 — Data and Model Poisoning

Severity: Critical

Training data or model weights corrupted with backdoors. In PCI scope, a poisoned model in a fraud detection or payment processing pipeline could be triggered to systematically approve fraudulent transactions or suppress fraud alerts. Req 6.5 (change management) and Req 11.3 (penetration testing) are the primary requirements.

PCI DSS v4.0 mapping

Requirement ID How it applies
Secure system changes Req 6.5.6 All model promotions tested for unexpected functionality before CDE deployment — poisoning as unexpected functionality
Audit log review Req 10.6.1 Automated monitoring of LLM outputs in CDE — poisoning indicators (systematic anomalous recommendations) detected
Penetration testing Req 11.3.1 Poisoning detection included in penetration testing programme for LLM applications in CDE
Targeted risk analysis Req 12.3.2 Targeted risk analysis for training pipeline security — documented controls and frequency of review

Mitigations for PCI DSS compliance

Foundational

  • Req 6.5.6: Test all model promotions for unexpected functionality — model integrity verification and output distribution analysis before CDE deployment
  • Document training data governance — source allowlisting, provenance, anomaly detection as Req 6.2 secure development practices
  • Establish model rollback capability — Req 6.5 change reversal requirement met by clean checkpoint procedure

Hardening

  • Req 10.6.1: Deploy automated monitoring of LLM outputs in CDE — poisoning indicators (systematic anomalous fraud scores, recommendation patterns) detected
  • Req 11.3.1: Include poisoning detection in penetration testing — backdoor trigger testing for CDE LLM models
  • Req 12.3.2: Document targeted risk analysis for training pipeline — threat actors, attack vectors, controls, review frequency

Advanced

  • Post-training backdoor detection as mandatory CDE deployment gate — Req 6.5 testing before production
  • Req 10.6.1: Integrate model output anomaly detection into PCI-scope SIEM — poisoning events treated as security incidents per Req 12.10
  • Conduct training pipeline adversarial testing — Req 11.3 penetration testing evidence

Tools

Tool Type Link
IBM Adversarial Robustness Toolbox Open-source https://github.com/Trusted-AI/adversarial-robustness-toolbox
CleanLab Open-source https://github.com/cleanlab/cleanlab

Cross-references

  • Agentic Top 10: ASI06 Memory and Context Poisoning
  • DSGAI 2026: DSGAI04 Data Model and Artifact Poisoning, DSGAI21 Disinformation via Data Poisoning
  • Other frameworks: NIST AI RMF MS-3.3 · MITRE ATLAS AML.T0070 · ISO 27001 A.8.27

LLM05 — Insecure Output Handling

Severity: High

LLM output passed to downstream systems without validation enables injection attacks. In PCI scope, output injection into payment processing logic could enable fraudulent transaction generation or CHD exfiltration. Req 6.2 (secure development) and Req 6.4 (public-facing application protection) apply.

PCI DSS v4.0 mapping

Requirement ID How it applies
Bespoke and custom software Req 6.2.4 Output handling in LLM integrations addresses all injection vulnerability classes — LLM output treated as untrusted
Public-facing web application protection Req 6.4.1 All interfaces consuming LLM output in PCI scope protected — input validation on all downstream consumers
Penetration testing Req 11.3.1 Output injection scenarios in penetration testing — SQL injection, command injection via LLM output
Audit log review Req 10.6.1 Monitoring for injection patterns in LLM output channels within CDE

Mitigations for PCI DSS compliance

Foundational

  • Req 6.2.4: Implement output validation as a secure development requirement — LLM output treated as untrusted input to all downstream CDE systems
  • Req 6.4.1: Apply application protection to all interfaces consuming LLM output in PCI scope — WAF or equivalent, schema validation
  • Never pass raw LLM output to payment processing logic — Req 6.2 secure design requirement

Hardening

  • Req 11.3.1: Include output injection in penetration testing scope — SQL injection, command injection via LLM output, results documented
  • Req 10.6.1: Monitor LLM output channels for injection patterns — automated log analysis covering CDE-facing output paths
  • Conduct DAST on all CDE interfaces consuming LLM output

Advanced

  • Deploy dedicated output security layer — Req 6.2 defence-in-depth evidence
  • Req 11.3.1: Adversarial output testing against specific payment logic consumers — document as CDE penetration testing evidence

Tools

Tool Type Link
OWASP ZAP Open-source https://www.zaproxy.org
Semgrep Open-source https://semgrep.dev

Cross-references

  • Agentic Top 10: ASI02 Tool Misuse, ASI05 Unexpected Code Execution
  • DSGAI 2026: DSGAI05 Data Integrity and Validation Failures
  • Other frameworks: OWASP ASVS V5 · CIS Controls CIS 16 · CWE-79

LLM06 — Excessive Agency

Severity: High

LLMs with excessive autonomy execute unintended actions. In PCI scope, excessive agency against payment systems — an agent that can initiate refunds, modify transaction records, or query cardholder tables autonomously — is a critical access control failure under Req 7 and Req 8.

PCI DSS v4.0 mapping

Requirement ID How it applies
Restrict access to system components Req 7.2.1 LLM tool access to CDE systems restricted to minimum required — read-only by default, write access formally approved
Access control system Req 7.3.1 Access control system enforces LLM tool scope — agent cannot exceed defined CDE access without explicit authorisation
Logging and monitoring Req 10.2.1 All LLM tool invocations in CDE logged — tool identity, parameters, invoking user identity, timestamp
Targeted risk analysis Req 12.3.2 Targeted risk analysis for LLM autonomous action scope in CDE — blast radius documented and accepted

Mitigations for PCI DSS compliance

Foundational

  • Req 7.2.1: Restrict LLM tool access to minimum required CDE access — read-only for analytics and advisory functions, write access requires formal approval and documented business justification
  • Req 7.3.1: Enforce LLM tool scope through access control system — agent cannot self-escalate beyond defined CDE permission scope
  • Require human confirmation for all LLM-initiated financial operations — Req 7 need-to-know requirement

Hardening

  • Req 10.2.1: Log all LLM tool invocations in CDE — tool identity, parameters, user identity, timestamp — immutable, meets Req 10 logging requirements
  • Req 12.3.2: Document targeted risk analysis for LLM autonomous actions — CDE systems accessible, operations possible, blast radius, controls implemented
  • Conduct quarterly access reviews for LLM tool permissions — Req 7 periodic review requirement

Advanced

  • Formally specify permitted LLM action scope for each CDE deployment — Req 7 documented least privilege
  • Req 11.3.1: Test excessive agency via injection in penetration testing — results documented as CDE security testing evidence

Tools

Tool Type Link
Guardrails AI Open-source https://github.com/guardrails-ai/guardrails
NeMo Guardrails Open-source https://github.com/NVIDIA/NeMo-Guardrails

Cross-references

  • Agentic Top 10: ASI01 Agent Goal Hijack, ASI02 Tool Misuse
  • DSGAI 2026: DSGAI06 Tool Plugin and Agent Data Exchange
  • Other frameworks: EU AI Act Art. 14 · ISO 42001 A.6.1.2 · AIUC-1 B006

LLM07 — System Prompt Leakage

Severity: High

System prompts containing security controls and internal logic are extracted. In PCI scope, system prompts may contain CDE topology, authentication logic, or processing rules that must be protected as sensitive configuration data under Req 3 and Req 7.

PCI DSS v4.0 mapping

Requirement ID How it applies
Protection of sensitive authentication data Req 3.5.1 System prompts containing CDE configuration treated as sensitive — encrypted at rest, access-controlled
Restrict access by need to know Req 7.2.1 System prompt access restricted — only authorised personnel with documented business need
Logging and monitoring Req 10.2.1 All access to system prompt storage logged — access to CDE configuration data requires audit trail
Bespoke software security Req 6.2.4 System prompt design prevents leakage — no cleartext CDE identifiers, tokens resolved at runtime

Mitigations for PCI DSS compliance

Foundational

  • Req 3.5.1: Classify system prompts containing CDE references as sensitive configuration — encrypted at rest, access-controlled, version-controlled
  • Req 7.2.1: Restrict system prompt access to authorised personnel with documented business need — access control configuration as Req 7 evidence
  • Remove all CHD, CDE topology, and authentication details from system prompts — use runtime token resolution per Req 6.2

Hardening

  • Req 10.2.1: Log all access to system prompt storage — access to CDE configuration data auditable per Req 10
  • Conduct prompt extraction testing — Req 11.3 security testing evidence
  • Req 6.2.4: Implement system prompt tokenisation — CDE identifiers replaced with opaque tokens

Advanced

  • Req 11.3.1: Include prompt extraction in penetration testing scope — CDE system prompt recovery tested
  • Deploy output classifier to block system prompt content in responses — Req 6.4 application protection

Cross-references

  • Agentic Top 10: ASI01 Agent Goal Hijack
  • DSGAI 2026: DSGAI15 Over-Broad Context Windows
  • Other frameworks: AIUC-1 B003/B009 · CWE-200 · ISO 27001 A.5.12

LLM08 — Vector and Embedding Weaknesses

Severity: Medium

Vector stores containing embeddings of CHD-adjacent documentation represent secondary attack surfaces that may be outside traditional PCI scope assessments but within CDE data exposure risk.

PCI DSS v4.0 mapping

Requirement ID How it applies
Protect stored account data Req 3.4.1 Embeddings derived from CHD-containing documents encrypted — PAN cannot be reconstructed from cleartext embeddings
Protect stored account data — SAD Req 3.5.1 Vector stores in CDE scope encrypted with strong cryptography
Restrict access Req 7.2.1 RBAC on all vector stores in CDE scope — no unauthenticated access
Penetration testing Req 11.3.1 Vector store attacks in CDE penetration testing — RBAC bypass, bulk extraction, embedding inversion

Mitigations for PCI DSS compliance

Foundational

  • Req 3.4.1: Ensure embeddings derived from CHD-containing documents cannot reproduce PANs — PAN masking before embedding generation prevents scope contamination
  • Req 7.2.1: Enable RBAC on all vector stores in PCI scope — no unauthenticated access in any environment
  • Req 3.5.1: Encrypt all vector stores in CDE scope — encryption at rest required

Hardening

  • Scope design: Apply PAN tokenisation before embedding generation — removes CHD from embedding scope, reducing PCI scope of vector infrastructure
  • Req 11.3.1: Include embedding inversion in penetration testing — validate CHD cannot be reconstructed, results documented as CDE testing evidence
  • Patch all vector database CVEs — Req 6.3 vulnerability management requirement

Advanced

  • Apply differential privacy to embeddings of CHD-adjacent content — Req 3 advanced data protection measure
  • Req 11.3.1: Conduct RBAC bypass testing for all vector stores in PCI scope — document as CDE testing evidence

Tools

Tool Type Link
Weaviate Open-source https://weaviate.io
ML Privacy Meter Open-source https://github.com/privacytrustlab/ml_privacy_meter

Cross-references

  • Agentic Top 10: ASI06 Memory and Context Poisoning
  • DSGAI 2026: DSGAI13 Vector Store Platform Security, DSGAI18 Inference and Data Reconstruction
  • Other frameworks: NIST AI RMF MS-2.5 · ISO 27001 A.8.3/A.8.24 · SOC 2 C2.1

LLM09 — Misinformation

Severity: Medium

LLMs generate incorrect content. In PCI scope, misinformation from an LLM providing payment security guidance, chargeback advice, or fraud detection recommendations could cause incorrect processing or compliance errors. Req 12.3 targeted risk analysis governs the acceptable risk tolerance for LLM advisory outputs in payment contexts.

PCI DSS v4.0 mapping

Requirement ID How it applies
Targeted risk analysis Req 12.3.2 Targeted risk analysis for LLM advisory outputs in payment context — what decisions are influenced, what is the consequence of incorrect output
Bespoke software security Req 6.2 LLM integration code specifies accuracy requirements — high-stakes payment outputs require verification before action
Audit log review Req 10.6.1 Monitoring for LLM accuracy degradation — output quality metrics reviewed
Security awareness programme Req 12.6.1 PCI staff trained on LLM output limitations — advisory status of AI recommendations communicated

Mitigations for PCI DSS compliance

Foundational

  • Req 12.3.2: Document targeted risk analysis for LLM advisory outputs in payment context — what payment decisions are influenced by LLM, acceptable accuracy threshold, human verification requirements
  • Req 12.6.1: Include LLM output limitations in PCI security awareness programme — staff handling CHD through LLM trained on advisory status
  • Human verification for all LLM-influenced payment decisions — Req 12.3 risk acceptance control

Hardening

  • Req 10.6.1: Implement production accuracy monitoring — LLM output quality reviewed as part of CDE log analysis programme
  • Deploy RAG grounded on authoritative, PCI-approved sources — Req 6.2 secure design for payment advisory
  • Req 12.3.2: Review and update targeted risk analysis when LLM model is updated — accuracy characteristics may change post fine-tuning

Advanced

  • Conduct domain-specific accuracy testing before CDE deployment — Req 6.5 testing before production change
  • Req 11.3.1: Include misinformation scenarios in penetration testing for payment advisory LLM — document as CDE testing evidence

Tools

Tool Type Link
TruLens Open-source https://github.com/truera/trulens
RAGAS Open-source https://github.com/explodinggradients/ragas

Cross-references

  • Agentic Top 10: ASI09 Human-Agent Trust Exploitation
  • DSGAI 2026: DSGAI21 Disinformation and Integrity Attacks
  • Other frameworks: EU AI Act Art. 13/50 · AIUC-1 F · NIST CSF 2.0 GV.OC-01

LLM10 — Unbounded Consumption

Severity: Medium

Adversarial inputs trigger disproportionate resource consumption. In PCI scope, availability of payment processing systems is a direct compliance requirement — resource exhaustion attacks against LLM applications in CDE are covered under Req 1 (network security) and Req 6.4 (application protection).

PCI DSS v4.0 mapping

Requirement ID How it applies
Network security controls Req 1.3.2 Network controls restrict inbound traffic to CDE — rate limiting at network layer for LLM applications
Public-facing application protection Req 6.4.1 LLM-powered public-facing applications protected against DoS — rate limiting and payload limits enforced
Audit log review Req 10.6.1 Automated monitoring for resource exhaustion patterns — unusual consumption volumes alerted
Targeted risk analysis Req 12.3.2 Targeted risk analysis for LLM availability in CDE — DoS impact on payment processing, controls documented

Mitigations for PCI DSS compliance

Foundational

  • Req 1.3.2: Implement network-layer rate limiting for CDE-facing LLM applications — traffic controls at the network boundary as Req 1 network security control
  • Req 6.4.1: Enforce rate limiting and payload limits on all public-facing LLM applications in PCI scope — application protection against resource exhaustion
  • Req 12.3.2: Document targeted risk analysis for LLM availability — DoS impact on payment processing, RTO/RPO, controls documented and accepted

Hardening

  • Req 10.6.1: Implement consumption anomaly monitoring — automated detection of unusual resource utilisation patterns in CDE LLM applications
  • Define BCP for LLM service failures affecting payment processing — Req 12 business continuity requirement
  • Per-user token limits and cost budgets — Req 6.4 application protection control

Advanced

  • Deploy sponge example detection — Req 6.4 advanced application protection for CDE
  • Req 11.3.1: Conduct adversarial load testing against CDE LLM applications — verify rate limiting holds, document as penetration testing evidence

Tools

Tool Type Link
LiteLLM Open-source https://github.com/BerriAI/litellm
Kong Gateway Open-source https://github.com/Kong/kong

Cross-references

  • Agentic Top 10: ASI08 Cascading Agent Failures
  • DSGAI 2026: DSGAI17 Data Availability and Resilience Failures
  • Other frameworks: ISA/IEC 62443 SR 7.6 (OT) · CIS Controls CIS 12 · NIST CSF 2.0 PR.IR-01

PCI DSS v4.0 compliance checklist for LLM applications

Scope determination (before anything else)

  • Determine whether LLM application is in CDE scope
  • Confirm no CHD in LLM training data (Req 3.3)
  • Confirm no SAD in any LLM data store (Req 3.3.1)
  • Map all data flows from LLM to CDE systems
  • Assess whether LLM infrastructure requires PCI segmentation

Data protection (Req 3, Req 4)

  • PANs in LLM scope encrypted with strong cryptography (Req 3.5.1)
  • PAN masking in all LLM outputs (Req 3.4.1)
  • SAD never stored in any LLM component (Req 3.3.1)
  • All CHD data flows encrypted in transit TLS 1.2 minimum (Req 4.2.1)
  • Embeddings of CHD-adjacent content cannot reproduce PANs (Req 3.4.1)

Secure development (Req 6)

  • LLM integration code addresses prompt injection as injection class (Req 6.2.4)
  • Public-facing LLM applications protected — WAF or equivalent (Req 6.4.1)
  • Model updates through change management with security testing (Req 6.5)
  • Output validation for all LLM outputs in CDE (Req 6.2.4)
  • System prompts contain no CHD or CDE identifiers (Req 6.2.4)

Access control (Req 7, Req 8)

  • LLM access to CDE systems restricted to minimum required (Req 7.2.1)
  • LLM tool access managed as need-to-know (Req 7.2.1)
  • Quarterly access reviews for LLM CDE permissions (Req 7)
  • LLM service accounts use unique identities (Req 8)
  • RBAC on all vector stores in CDE scope (Req 7.2.1)

Logging and monitoring (Req 10)

  • All LLM access to CDE logged with user identity (Req 10.2.1)
  • All LLM tool invocations in CDE logged (Req 10.2.1)
  • Automated log analysis covers LLM-specific anomalies (Req 10.6.1)
  • System prompt access logged (Req 10.2.1)

Testing (Req 11)

  • Prompt injection in penetration testing scope (Req 11.3.1)
  • Output injection in penetration testing scope (Req 11.3.1)
  • Excessive agency tested for CDE-facing agents (Req 11.3.1)
  • Vector store attacks in penetration testing (Req 11.3.1)

Vendor and policy (Req 12)

  • LLM component vendors with CDE access in TPSP list (Req 12.8.1)
  • Written agreements with LLM vendors (Req 12.8.3)
  • TPSP compliance status monitored for LLM vendors (Req 12.8.4)
  • Targeted risk analyses for all LLM-specific risks (Req 12.3.2)
  • Security awareness programme covers LLM limitations (Req 12.6.1)

Implementation priority

Phase LLM entries PCI DSS requirements Rationale
1 — Scope first LLM02 Req 3, Req 4 CHD scope determination and data protection are the foundation — everything else depends on this
2 — Access and logging LLM06, LLM07 Req 7, Req 10 Access control and audit logging are core PCI requirements
3 — Secure development LLM01, LLM05 Req 6.2, Req 6.4 Injection and output security as secure development requirements
4 — Vendor and testing LLM03, LLM04 Req 12.8, Req 11.3 TPSP programme and penetration testing programme
5 — Ongoing LLM08-LLM10 Req 12.3, Req 6.5 Risk analysis, embedding scope, availability

See also


References


Changelog

Date Version Change Author
2026-03-25 2026-Q1 Initial mapping — LLM01-LLM10 full entries with PCI DSS v4.0 compliance checklist OWASP GenAI Data Security Initiative

Maintained by the OWASP GenAI Data Security Initiative. Part of the GenAI Security Crosswalk: https://github.com/emmanuelgjr/GenAI-Security-Crosswalk