What does an AI governance framework need from the security team?
Most of the evidence in the template's right-hand column is security evidence. Policy and impact assessment sit with risk, legal and the business. Inventory, logging, oversight, testing, deactivation and incidents are security operations work, and they are where audits fail.
Three things make this harder with AI agents than with earlier AI.
Agents are identities. An agent that can read data, call tools and change systems is a privileged account. It needs an owner, least privilege and a reviewed scope, the same as any service account.
The evidence is spread across systems. A single agent action can involve a model provider's audit log, an identity provider's sign-in, a tool call, a network connection and a change in a target system. Each provider audits its own part well. None of them joins the chain. Raji and colleagues' work on internal algorithmic auditing makes the same point about AI generally: once a system is deployed, emergent issues can be difficult or impossible to trace back to their source. Their answer is an internal audit process that produces documentation throughout development, so the record exists before it is needed.
Data can carry instructions. Agents that read email, documents or web pages can be steered by text planted in them, a risk shown in research on indirect prompt injection and ranked first in the OWASP Top 10 for LLM applications. The testing row of the template has to include it.
MITRE's ATLAS knowledge base catalogues attack techniques against AI systems and is a useful input to the risk assessment row.
Where SenseOn fits, and where it does not
SenseOn does not write your AI policy, run your impact assessments or certify you against ISO/IEC 42001. Those belong to your risk, legal and compliance owners and your certification body.
SenseOn supplies evidence for the security-heavy rows of the template:
- AI inventory (GOVERN 1.6; A.4.2). The 7-day AI inventory produces a named, human-reviewed list of the AI apps, agents and tools found in your estate: who owns them, what they can reach, and which actions lack an approval gate.
- Logging and monitoring (A.6.2.8; MANAGE 4.1). Security for AI connects provider, identity, tool, network and target-system evidence into one chain, so you can reconstruct what an AI workflow did.
- Deactivation and reversal (MANAGE 2.4). The same chain shows what access to contain and whether a change was reversed.
- Governance of SenseOn's own agents. Every AI and analyst action in SenseOn writes to Decision Trace, an append-only record with 100% coverage by design. SenseOn is certified to ISO/IEC 27001:2022 by British Assessment Bureau (UKAS-accredited).
Related reading: agentic AI security covers the controls for AI agents in more depth, the board AI risk assessment sets out the decisions a board has to own, shadow AI covers the unsanctioned tools your inventory will find, and the agentic SOC guide applies the same principles to AI inside security operations.
Frequently asked questions
What is an AI governance framework?
An AI governance framework is the set of roles, policies, processes and evidence an organisation uses to decide which AI systems it runs, what each may do, and how it proves they behaved as approved. Most organisations now build one on ISO/IEC 42001 for structure and the NIST AI Risk Management Framework for risk method, with the EU AI Act as a legal overlay where it applies.
What is the difference between ISO 42001 and NIST AI RMF?
ISO/IEC 42001 is a certifiable management-system standard: it sets requirements an auditor tests, in the same shape as ISO/IEC 27001, plus 38 reference controls in Annex A. The NIST AI RMF is a voluntary risk framework organised into Govern, Map, Measure and Manage. ISO tells you what system to run; NIST tells you how to assess and treat the risks inside it. NIST publishes an official crosswalk between them.
Is ISO 42001 certification mandatory?
No. ISO/IEC 42001 is voluntary. Organisations pursue certification because customers, regulators and procurement teams increasingly ask for evidence of AI governance, and a certificate from an accredited body is a recognised way to show it.
Does ISO 42001 help with EU AI Act compliance?
It helps, but it is not the same thing. The EU AI Act sets legal obligations that depend on each system's risk category, and ISO/IEC 42001 does not on its own demonstrate conformity with them. A working ISO/IEC 42001 management system does give you much of the structure the Act expects, such as risk management, documentation, logging and human oversight, so the evidence can be reused.
What should an AI governance framework include?
At minimum: an AI policy, named owners for every system, an inventory of AI in use, impact and risk assessments, data provenance records, a statement of what each system may do unattended, security testing, logs that link AI actions to identities and target systems, a way to disable and reverse, supplier controls, an incident process, and scheduled audit and review. The mapping template in this guide covers each one.
Sources
How SenseOn writes, checks and corrects its content: editorial standards.
