FoxPointe Security Hub

Abstract

Agentic AI is rapidly moving from concept to enterprise adoption, bringing new risks that traditional compliance frameworks were not designed to address. ISO/IEC 42001, the world’s first AI governance standard, is emerging as the go-to framework for organizations that want to manage AI responsibly. At the same time, PCI DSS v4.0.1 and ISO 27001 remain critical foundations, providing proven controls for authentication, logging, and non-human identities. In this article, we explore how businesses can shift left on AI governance, align to emerging standards, and stay secure and compliant in the era of Agentic AI.

A New Era of AI Risk

Artificial Intelligence has been a headline topic for years, but 2024 marked a turning point. Generative AI tools such as ChatGPT, Copilot, and Gemini became everyday business utilities. Now, a new wave of innovation is emerging in the form of Agentic AI. Unlike traditional chatbots, these systems do more than generate text or images; they act as autonomous agents that can plan, decide, and execute workflows across enterprise environments.

This evolution brings both opportunity and risk. During Cybersecurity Awareness Month, it is worth asking: are existing compliance frameworks such as PCI DSS, ISO standards, and others prepared for Agentic AI? Or are businesses moving ahead faster than regulation and governance can keep up?

Understanding the Terminology

After attending several AI security summits hosted by the Cloud Security Alliance, one thing became clear: the vocabulary around AI is expanding almost as quickly as the technology itself. Below are some key terms worth understanding.

• Generative AI – AI that creates new content such as text, images, or code.
• Agentic AI – autonomous AI agents that can pursue goals, make decisions, and interact with systems without constant human oversight.
• Large Language Models (LLMs) – the foundational models powering most generative and agentic systems.
• Non-Human Identities (NHIs) – privileged accounts such as API keys, service accounts, and OAuth tokens. These resemble privileged user accounts under PCI DSS and ISO frameworks, and managing them securely is a core part of AI governance.
• Deterministic AI – systems or algorithms that always produce the same output from the same input, such as traditional rule-based software or a Python script. Deterministic approaches are predictable and ideal for tasks that demand consistency.
• Non-Deterministic AI – models like LLMs that may produce different outputs from the same prompt due to probabilistic reasoning. This variability enables creativity and contextual adaptation, but it also introduces unpredictability that must be managed.
• Vibe Coding – programming through natural language intent rather than strict syntax.
• Hallucinations – when AI outputs confident but false or misleading information.
• MCP Servers (Model Context Protocol) – infrastructure that connects AI agents to tools and environments.
• Guardrails – controls that limit or constrain AI behavior to prevent misuse.

For business and compliance leaders, understanding these terms is the first step toward understanding how risks are changing.

The Risks of Agentic AI

Agentic AI introduces a very different risk profile compared to earlier automation tools:

• Prompt Injections – AI agents can be manipulated into executing harmful commands such as deleting files or sending credentials. Unlike a human, an agent may not recognize malicious intent.
• Jailbreaks – attackers can bypass built-in safety guardrails through techniques like escalating requests or reframing context. This resembles privilege escalation or injection attacks in traditional systems.
• Fragmented Ecosystem of Models – hundreds of LLMs are now available, each with different training data, safety controls, and governance. Independent initiatives such as RiskRubric.AI are starting to rank models, but the landscape remains highly variable.
• Hallucinations and Reasoning Failures – the tendency of LLMs to generate false or nonsensical outputs remains a major concern. In 2023, several U.S. attorneys were sanctioned after submitting legal briefs that cited fabricated case law generated by AI. Even when not hallucinating, models can reason incorrectly when users introduce doubt, such as asking, “Are you sure about that?” These risks highlight the need for human oversight.
• Model Autonomy Risks – an AI agent that chains multiple tasks can generate unintended outcomes. A single misconfigured instruction could affect production environments at scale.
• Shadow AI – similar to Shadow IT, employees experimenting with unapproved AI tools risk leaking sensitive data, intellectual property, or regulated information.
• Compliance Gaps – many regulatory frameworks are still built around human logins, static access rights, and server configurations. They rarely address non-human identities, autonomous decision-making, or the dynamic behavior of AI agents.

In short, Agentic AI creates new forms of unpredictability and potential misuse that organizations must learn to govern.

Frameworks Trying to Keep Pace

The good news is that standards bodies and regulators are moving quickly to close the gap. Notable developments include:

• CSA AI Control Matrix – expands the Cloud Controls Matrix with AI-specific safeguards.
• NIST AI RMF – focuses on trustworthy, explainable, and accountable AI.
• EU AI Act – classifies AI systems into risk tiers with corresponding compliance obligations.
• EU NIS2 Directive – broadens cybersecurity governance obligations across Europe, indirectly influencing AI deployments.
• BSI AI C4 – Germany’s detailed criteria for trustworthy AI, offering a glimpse into more prescriptive regulation ahead.
• ISO/IEC 42001:2023 – the world’s first international AI governance standard, rapidly gaining traction. ISO 42001 establishes a structured management system for AI, much like ISO 27001 transformed information security. For most organizations, ISO 42001 will become the central governance framework moving forward.
• PCI DSS v4.0.1 – while not AI-specific, PCI DSS was intentionally written with flexible, principle-based language that helps future-proof its requirements. Controls around authentication, access management, logging, and non-human identities apply directly to AI agents. For payment environments, PCI DSS provides a strong baseline that can be extended with ISO 42001 or NIST’s AI RMF. The PCI Security Standards Council has also released initial guidance on the use of AI in payment environments.

Despite these advances, the frameworks are still catching up to the speed of AI adoption. Companies moving faster than regulators will need to bridge the gap through internal governance and accountability.

What This Means for Businesses

Organizations adopting Agentic AI early face two parallel challenges: leveraging its benefits while proving compliance to auditors, regulators, and customers.

Key implications include:

• Strategic Adoption – companies are deploying AI for secure coding, compliance monitoring, fraud detection, and customer support. Each use case carries distinct risks that must be identified and mitigated.
• Shift-Left Governance – just as DevSecOps moved security earlier in the lifecycle, AI governance must also shift left. Embedding risk assessments, red-team testing, and explainability reviews during the design phase prevents rework and audit findings later.
• GRC Alignment – mapping AI practices to ISO/IEC 42001, NIST AI RMF, and CSA AI Control Matrix provides a defensible compliance posture. For organizations in the payments space, PCI DSS remains foundational and integrates naturally with these broader frameworks.
• Demonstrating Compliance – auditors will expect tangible evidence: logs of AI agent actions, risk assessments, non-human identity reviews, and compensating controls where frameworks have gaps.

Enterprises that fail to adopt a structured governance model risk falling out of compliance or suffering an AI-driven incident that existing frameworks were not designed to address.

The Path Forward: Awareness and Action

So, are compliance frameworks ready for Agentic AI? Not entirely. They are evolving, but the pace of AI innovation continues to exceed regulatory change.

Rather than waiting for new rules, organizations should take proactive steps now. Cybersecurity Awareness Month is an ideal time to assess readiness and raise awareness of how quickly AI is reshaping the threat landscape.

Forward-looking organizations should:

• Adopt ISO/IEC 42001 or map to NIST AI RMF to establish governance early.
• Treat AI agents as non-human identities under frameworks such as PCI DSS.
• Invest in guardrails, monitoring, and red-teaming for AI-driven processes.
• Educate leadership teams on both the strategic opportunities and the compliance risks introduced by Agentic AI.

Building Readiness Together

Agentic AI has the power to transform business operations and decision-making, but it also challenges compliance frameworks that were never built for autonomous systems. While some prescriptive controls now seem dated, PCI DSS and ISO standards still offer strong, practical foundations that can anchor organizations until new AI-specific frameworks mature.

The path forward is clear: adopt governance early, align with emerging standards such as ISO/IEC 42001, and manage AI risk with the same rigor used for other critical systems.

At FoxPointe Solutions, we are already helping clients close this gap through PCI DSS assessments, ISO 27001 implementations, and ISO/IEC 42001 consulting and internal audits. As Agentic AI becomes part of everyday business, we can help ensure your compliance posture remains strong, adaptable, and trusted.