Read Time: 5 minutes
Organizations cannot succeed in securing AI using legacy models alone because AI technologies present a broader attack surface and introduce unique threats, such as prompt injection, data poisoning, and model theft or extraction.
Defining Intent
Selecting the right framework begins with defining organizational intent. Cybersecurity leaders must determine if their primary driver is:
- Regulatory Compliance: Meeting specific legal mandates.
- Formal Certification: Seeking internationally recognized benchmarks like ISO/IEC 42001 to demonstrate auditability.
- Technical Controls: Requiring granular guidance for specific technical implementation.
This intent must then be matched to specific AI deployment patterns, ranging from traditional and generative AI to fully autonomous, multiagent architectures, to ensure the framework fits the organization's technical complexity.
The Hybrid Imperative
Relying on a single framework almost guarantees security gaps because different models focus on distinct areas, such as technical controls, organizational policy, or auditability. Forward-thinking organizations adopt a hybrid approach, blending the strengths of multiple standards:
- ISO/IEC 42001: Best for comprehensive AI governance, life cycle oversight, and formal certification.
- NIST AI RMF (100-1): Ideal for establishing governance and strategic priorities, and defining trustworthiness characteristics such as being explainable, resilient, and privacy-enhanced.
- OWASP AI Frameworks: Provides tactical, actionable guidance focused on identifying exploitable vulnerabilities and adversarial testing.
- Vendor-Specific Deployment: Securing AI within a defined vendor ecosystem using frameworks like Google's Secure AI Framework or Microsoft's Responsible AI Framework.
Frameworks Must Evolve
Cybersecurity leaders must standardize on their framework rapidly, so guidance is available before teams start new projects. Early implementation enables AI teams to move faster and prevents late, costly project delays caused by risks that were not identified early enough. Because the AI field is evolving faster, organizations should reassess at least annually, or whenever specific triggers occur:
- New AI use cases are introduced
- The regulatory environment shifts or new requirements are passed
- Organizational AI objectives or risk tolerance change
Starting from Within
It is critical to understand that AI-specific frameworks do not replace existing application security or cybersecurity risk frameworks; they must be implemented in addition to them. Crowe Indonesia Teknologi helps organization to have effective adoption begins by addressing the cybersecurity debt exacerbated by AI and strengthening existing controls before layering on AI-specific measures.