The security terms look technical.  
The gaps behind them are governance.

Week four of MIT Applied Agentic AI for Organisational Transformation moved into security architecture.

Programme governance has to absorb every term on this list — before deployment, not after.

Fourteen terms. Half AI-native. Half classic cybersecurity. Every single one carries a governance owner — or it doesn't.

That absence is the story.

Take Phased Deployment.

The definition is clean — AI is introduced gradually, starting with low-risk applications before scaling to higher-stakes ones.

But the question that matters isn't technical. Who defines "low-risk"? Who owns the criteria for moving from phase one to phase two? Who holds the authority to halt progression when the signals don't support it?

If the phase gates don't have named owners and documented criteria, "phased" is a narrative the vendor tells the steering committee.

Sandboxing sounds like a familiar containment principle. Then the pressure to ship arrives.

Sandboxing isolates an agent from the wider system. So the governance question becomes — who authorises release from sandbox to production? What evidence is required? Who carries the accountability when sandboxed behaviour doesn't replicate in the live environment?

Most sandbox release decisions get made by whoever is in the room. That's not governance. That's scheduling.

Deepfake looks like an external threat. It's also a control failure waiting to surface.

When a CFO's cloned voice authorises a wire transfer, who owns the verification chain? Is there a documented second-channel check for high-value instructions? Who signed that protocol off?

Most firms can describe their financial controls. Few can describe their authenticity controls.

The pattern I see consistently in transformation programmes evaluating agentic AI:

→ Cybersecurity teams own the controls.  
→ Governance teams don't see the AI-specific attack surface.  
→ The gap between them is where regulators will arrive first.

AI security at the leadership level isn't about understanding the attacks.

It's about assigning ownership before the agent is deployed. Who authorises phase progression? Who signs off on the sandbox release? Who owns the verification chain when deepfakes, spoofing, and impersonation converge?

These aren't CISO questions alone. They're governance ones.

Week five next.

🔔 Follow Justin R. for more insights on AI enterprise transformation
♻️ Save and share this with anyone signing off agent deployment in a regulated environment
🚀 Subscribe to The Transformation Constant — my newsletter