Skip to content

AI security risks, mapped to PCI DSS, ISO 27001, NIST CSF and HIPAA

By Sam Rivera, Founder, SentinelPanda · November 7, 2026 · 6 min read · AI Governance

For most companies AI did not arrive with a compliance regime of its own. It arrived inside the regimes they already have, and each of those already has a requirement that applies.

The risks are familiar; the carrier is new

Strip away the novelty and the security risks of AI in a regulated business are a short list. Sensitive data is pasted into a prompt and leaves the organisation. A tool nobody approved is in daily use. An integration holds more access than it needs. A vendor processes data under terms nobody read. Actions are taken that nobody logged. Content an attacker controls ends up steering a system that can act.

None of these is new to a security team. What is new is how fast AI tools are adopted, and how many places data can now flow without a project, a ticket or a review. That is why the frameworks you already follow are the right place to start: their requirements apply to an AI tool exactly as they apply to any other system component, supplier or data flow.

1. Sensitive data in prompts

This is the most common AI incident in practice, and the least dramatic. Someone pastes a customer record, a support transcript or a database extract into a chat assistant to save ten minutes.

  • PCI DSS v4.0.1: account data belongs only in systems inside your documented scope (12.5.2); sensitive authentication data must not be kept after authorisation (3.3.1); a vendor that receives account data is a third-party service provider under 12.8.
  • ISO/IEC 27001:2022: classification and transfer rules (A.5.12, A.5.14), data leakage prevention (A.8.12) and the cloud services control (A.5.23) all reach a hosted AI tool.
  • NIST CSF 2.0: protection of data at rest, in transit and in use (PR.DS-01, PR.DS-02, PR.DS-10), and a data inventory that shows where sensitive data goes (ID.AM-07).
  • HIPAA: sending PHI to a service that has not signed a business associate agreement is a disclosure the rules do not permit, and the minimum-necessary standard limits what is sent even when an agreement exists.

2. Shadow AI

You cannot assess a tool you do not know about. Browser extensions, meeting transcribers, coding assistants and AI features switched on inside existing SaaS products are the usual inventory gap.

  • PCI DSS: an inventory of in-scope system components (12.5.1), and of the third-party components built into your own software, such as AI libraries and SDKs (6.3.2).
  • ISO 27001: the asset inventory (A.5.9) and acceptable-use rules (A.5.10).
  • NIST CSF: inventories of software and services (ID.AM-02) and of services provided by suppliers (ID.AM-04), backed by policy (GV.PO).
  • HIPAA: the risk analysis (164.308(a)(1)(ii)(A)) covers every system that creates, receives, maintains or transmits ePHI, including the ones IT did not buy.

3. Prompt injection and unsafe output

When an AI system reads content an attacker can influence (an email, a ticket, a web page, a document), that content can carry instructions. If the system can act on its output, the attacker can act through it. The control is mostly secure engineering: treat model output as untrusted input, validate it, and limit what it can trigger.

  • PCI DSS: software engineering techniques that prevent common attacks, including injection (6.2.4), and protection of public-facing web applications (6.4.1, 6.4.2).
  • ISO 27001: application security requirements (A.8.26), secure coding (A.8.28) and security testing in development and acceptance (A.8.29).
  • NIST CSF: secure software development practices built into the life cycle (PR.PS-06).
  • HIPAA: integrity controls that protect ePHI from improper alteration (164.312(c)(1)).

4. Over-privileged agents and integrations

An assistant connected to a mailbox, a CRM or a payment console can do whatever its credential allows. Shared service accounts and long-lived API keys make that reach both larger and harder to trace.

  • PCI DSS: least privilege by role (7.2.1, 7.2.2), and management of system and application accounts and their credentials (8.6.1 to 8.6.3).
  • ISO 27001: access control and access rights (A.5.15, A.5.18), privileged access rights (A.8.2) and information access restriction (A.8.3).
  • NIST CSF: access permissions managed with least privilege and separation of duties (PR.AA-05).
  • HIPAA: information access management (164.308(a)(4)) and technical access control (164.312(a)(1)).

5. AI vendors and the model supply chain

Most organisations consume AI as a service. The questions that matter are ordinary supplier questions, asked precisely: what is retained, for how long, where, who can see it, and whether any of it is used to train models.

  • PCI DSS: third-party service provider management, including written acknowledgement of responsibility and a record of who manages which requirement (12.8.1 to 12.8.5).
  • ISO 27001: supplier relationships, supplier agreements, the ICT supply chain and supplier monitoring (A.5.19 to A.5.22), plus cloud services (A.5.23).
  • NIST CSF: the cybersecurity supply chain risk management category (GV.SC).
  • HIPAA: a business associate agreement before PHI is shared (164.308(b)(1), 164.314(a)), with the same obligations passed down to the vendor's subcontractors.

6. No record of what the AI did

When an AI-assisted action goes wrong (a refund issued, a record changed, a message sent), the first question is who did it. "The assistant" is not an answer an auditor accepts.

  • PCI DSS: audit logs that capture individual access to cardholder data (10.2.1, 10.2.1.1), and regular review of them (10.4).
  • ISO 27001: logging (A.8.15) and monitoring activities (A.8.16).
  • NIST CSF: continuous monitoring (DE.CM) and adverse event analysis (DE.AE).
  • HIPAA: audit controls that record and examine activity in systems containing ePHI (164.312(b)).

7. AI-made social engineering

Voice cloning and fluent written pretexts have made the old advice (listen for odd phrasing) far less useful. The defence is process: call-back verification for payment changes and access requests, whatever the voice sounds like.

  • PCI DSS: mechanisms that detect and protect against phishing (5.4.1), and security awareness covering phishing and social engineering (12.6.3.1).
  • ISO 27001: information security awareness, education and training (A.6.3).
  • NIST CSF: awareness and training (PR.AT).
  • HIPAA: security awareness and training (164.308(a)(5)).

Where the four frameworks stop

None of these frameworks evaluates the model itself: whether its output is accurate, whether it is biased, whether a person reviews the decisions it informs. Those are AI management questions, covered by ISO/IEC 42001 and the NIST AI Risk Management Framework, and in the EU by the AI Act. The security frameworks protect the data and systems around the model; the AI frameworks govern the model's behaviour and use. Most organisations need some of both, and the overlap is larger than it looks.

One control set, four frameworks

The efficient approach is to write one set of AI controls (an approved-tools list, a data rule for prompts, distinct identities for agents, vendor terms on retention and training, logging of AI actions, call-back verification) and record once which requirement in each framework every control satisfies. Evidence collected for one then counts for all four.

That is how SentinelPanda models it: one control, mapped across frameworks, with evidence attached once and reused wherever the mapping applies.

The short version

  • Keep an inventory of AI tools, including AI features inside software you already use.
  • Write down what data may go into a prompt, and what may not.
  • Give every agent and integration its own identity and the smallest grant that works.
  • Treat AI vendors as suppliers: retention, location, access and training use, in writing.
  • Log what AI-assisted systems do, attributable to an identity.
  • Verify payment and access changes out of band, whatever the caller sounds like.
When an AI agent enters PCI scope AI tools and PHI under HIPAA ISO 27001 controls for AI NIST CSF 2.0 for AI systems

Run your compliance program in one workspace.