COMMENTARY: Every major cloud and data platform now publishes guidance on how organizations should secure artificial intelligence (AI).From Google’s Secure AI Framework (SAIF) and Amazon Web Services’ CAF-AI and Generative AI best practices to Microsoft’s Secure AI and Responsible AI standards, enterprises are surrounded by frameworks.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]Add in guidance from NIST, the Cloud Security Alliance, and OWASP, and we have a rich, but often overwhelming landscape of AI security blueprints.These frameworks are valuable. They define principles, highlight risks, and establish a common language between security, engineering, and leadership teams. But many organizations struggle to move beyond documentation and compliance checklists. They find that frameworks describe what good looks like, but they rarely explain how to enforce trust when AI systems operate dynamically at runtime.Too often, organizations try to “pick the winner” among AI security frameworks. In practice, strong programs integrate multiple perspectives. Lifecycle-oriented frameworks offer structure, governance models clarify accountability, and platform-specific guidance informs implementation details.Enterprises that blend these strengths gain flexibility, especially in multi-cloud and hybrid environments, while avoiding lock-in to a single interpretation of AI risk.By defining just two or three anchors, enterprises create a consistent lens for evaluating tools, processes, and designs, regardless of how frameworks evolve.Metrics transform AI security from an abstract concept into an operational discipline.No single framework or vendor can secure AI end-to-end. Frameworks show direction, but controls deliver outcomes. Organizations that succeed will integrate guidance across frameworks, define clear security anchors, and invest in control planes that enforce trust continuously, at runtime, where AI risk actually lives. That shift can turn AI security from aspiration into actual execution.Nauman Mustafa, Chief Strategy Officer & VP of Solutions Engineering, BritiveSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
From governance to runtime reality
Most AI security programs stall at policy. Risk assessments are written, committees are formed, and controls are mapped on paper. Meanwhile, AI workloads, models, agents, pipelines, and APIs, are making decisions in real-time, interacting with sensitive data and infrastructure.This gap highlights a critical shift enterprises must make: a move from governance-first security to operational trust. We need to treat frameworks as roadmaps, not endpoints. We have to translate them into enforceable controls that operate continuously across the AI lifecycle.Better to think of them as an AI trust control plane: a unifying layer that turns identity, data, and model governance into real-time enforcement rather than static assurance.Map frameworks to the AI lifecycle
Despite their differences, major AI security frameworks align around a common lifecycle: Data ingestion; model training, validation, deployment, and monitoring.When frameworks are mapped to this lifecycle, they stop being abstract principles and start informing control architecture. Instead of asking whether a framework is “implemented,” organizations can ask a more practical question:What controls exist at each lifecycle stage, and how are they enforced?Five control domains for AI frameworks
Across frameworks, five recurring control domains consistently determine whether AI security programs succeed or fail:- Identity and access governance: AI systems introduce a surge of non-human identities. These identities should never have static or overly permissive access. Enforcing just-in-time, time-bound authorization for both humans and machines has become foundational to operational trust.
- Data provenance and lineage: Trustworthy AI depends on trustworthy data. Teams need to trace data from its source through training, fine-tuning, and inference. Verified lineage reduces the risk of poisoning, misuse, and regulatory exposure.
- Model integrity and security: Treat models as critical software assets. This includes scanning for vulnerabilities, testing against adversarial inputs, validating supply chains, and monitoring for anomalous behavior.
- Runtime authorization and enforcement: Extend zero-trust principles into AI execution environments. Access decisions should consider context, who or what requests access, for how long, and under what conditions, rather than relying on standing permissions.
- Governance and oversight: Accountability cannot sit in one team. Effective programs establish shared ownership across security, engineering, data, and compliance, with measurable outcomes rather than symbolic oversight.
Establish non-negotiable security anchors
Frameworks define what to protect, and controls define how. But many mature organizations add a third layer: security anchors. These are a small set of non-negotiable principles that guide every architectural and operational decision.Examples include:- AI agents and services must never hold standing privileged access.
- All privileged access must be contextual, time-bound, and auditable.
- Data used for training and inference must have verified provenance.
Measure what matters
We also must make operation trust measurable. Useful indicators include:- Percentage of AI identities using just-in-time access.
- Coverage of verified data lineage.
- Mean time to revoke privileged access.
- Percentage of models with signed metadata.
- Composite AI risk or trust maturity scores.
