Reflections of a Concerned Practitioner

As a kid I was a big fan of the famous TV series The Time Tunnel with the notorious couple Dr. Doug Phillips (played by Robert Colbert) and Dr. Tony Newman (played by James Darren) – I was literally stuck to the screen every summer. I also remember that there’s a scene in almost every time-travel movie where somebody says:

“We changed one small thing… and now everything is broken.”

That line stayed with me recently while watching the PocketOS incident unfold.

For those who missed it, the CEO of PocketOS, a software company servicing the car rental industry, publicly described how an AI agent reportedly wiped production data and backups within seconds. Even more surreal, Claude later acknowledged it had effectively “jumped ahead of itself.” The damage, however, was already done.

That single moment may end up being remembered as one of the first mainstream reminders that agentic AI is not behaving like traditional software.

It behaves more like a highly capable junior employee with:

• unlimited confidence,
• admin access,
• incredible speed,
• and absolutely no fear.

What could possibly go wrong?

That combination should concern every CISO, architect, regulator and executive.

AI Is Turning Us Into Modern Time Travellers

AI gives humanity something extraordinary. For the first time in history, we can meaningfully simulate the future at scale.

Not literally, of course. Nobody is stepping into a DeLorean.

But modern AI systems allow us to:

• predict adversarial behaviour,
• identify emerging fraud patterns,
• model cyber attack paths,
• simulate geopolitical impacts,
• forecast operational failures,
• optimise supply chains,
• and accelerate scientific discovery.

In cyber security specifically, AI allows defenders to compress time:

• detect threats earlier,
• predict attacker movement,
• correlate telemetry faster,
• and outpace adversaries before damage occurs.

That is incredibly powerful.

Used correctly and ethically, AI can help us become modern time travellers, seeing signals earlier and acting before events fully unfold.

The problem?

Many organisations are building systems capable of accelerating decisions faster than they can govern them. That is where things start becoming dangerous

The Agentic AI Gold Rush

The market is currently experiencing what feels like the early cloud era all over again and everyone following the hype.

Every vendor suddenly has:

• AI agents,
• autonomous copilots,
• orchestration frameworks,
• AI assistants,
• autonomous workflows,
• and “fully agentic platforms.”

The pressure from boards and investors is enormous:

• reduce costs,
• increase automation,
• improve customer experience,
• eliminate manual operations,
• and accelerate scale.

The result is predictable: organisations are racing to operationalise AI before governance frameworks have fully matured.

Sound familiar?

We did something similar with cloud adoption:

• move fast,
• connect everything,
• figure out governance later.

Many organisations are now repeating the same pattern with agentic AI, except this time the systems are not static infrastructure.

They think. They reason. They take action. They chain decisions together. They interact with APIs, datasets, workflows and external agents autonomously and increasingly, they can modify environments without human approval.

That changes everything

From AI Assistant to Autonomous Actor

Traditional software follows deterministic logic. You click a button. The system executes a known function. Agentic AI changes the model entirely.

Modern agents:

• interpret intent,
• build execution plans,
• orchestrate tasks,
• call APIs dynamically,
• interact with other agents,
• and continuously adapt behaviour based on context.

This creates enormous opportunity. It also creates volatility.

The real problem is not hallucination alone.The real problem is autonomous systems taking legitimate actions based on flawed reasoning, incomplete context, or misunderstood objectives.

That distinction matters. Hallucination inside a chatbot is embarrassing.

Hallucination inside:

• cloud infrastructure,
• financial systems,
• CI/CD pipelines,
• healthcare workflows,
• identity platforms,
• or production environments

becomes operationally catastrophic.

The PocketOS incident is important because it exposed something the market has been quietly ignoring:

Agentic systems can act faster than humans can intervene.

The Missing Layer: The AI Control Plane

This is why Gartner and many others are now talking about the emergence of the AI Control Plane. Personally, I believe this becomes one of the most important architectural shifts of the next 3–5 years.

Most organisations today have:

• cloud control planes,
• identity control planes,
• network control planes,
• orchestration layers,
• security management platforms.

But very few have a meaningful way to:

• monitor AI behaviour,
• govern autonomous decisions,
• enforce policy boundaries,
• validate outcomes,
• or audit agentic activity across environments.

That gap is becoming increasingly dangerous.

The AI Control Plane is emerging as the connective governance layer between:

• AI models,
• agents,
• APIs,
• orchestration frameworks,
• data flows,
• and operational systems.

Think of it as: part security layer, part governance layer, part operational oversight platform. It revolves around three foundational pillars:

1. Auditability – “What exactly did the AI do?”

This sounds simple. It isn’t.

Most organisations currently have very limited visibility into:

• why agents made decisions,
• which datasets influenced outcomes,
• what APIs were called,
• what chain-of-thought triggered actions,
• or how agents interacted with other agents.

That becomes a massive issue in:

• regulated industries,
• incident response,
• legal disputes,
• fraud investigations,
• and cyber events.

Imagine trying to investigate:

• why an AI deleted cloud resources,
• approved a fraudulent transaction,
• exposed sensitive data,
• or altered production configurations

without a reliable audit trail.

That is where the market currently is in many cases.

The AI Control Plane introduces:

• decision traceability,
• workflow telemetry,
• interaction logging,
• identity attribution,
• and behavioural observability.

Without auditability, AI becomes operationally opaque.

Opaque systems eventually become unmanageable systems.

2. Governance – “What is the AI allowed to do?”

This is where things become extremely interesting. Many organisations still think AI governance means:

• policies,
• ethics committees,
• awareness training,
• or PowerPoint frameworks.

Those matter. But operational governance is different.

Operational governance means:

• defining permissions,
• enforcing boundaries,
• limiting execution paths,
• restricting agent authority,
• controlling data access,
• validating integrations,
• and implementing runtime guardrails.

In other words:

AI governance is increasingly becoming an engineering discipline.

This is where technologies like:

• MCP (Model Context Protocol),
• A2A (Agent-to-Agent architectures),
• secure orchestration frameworks,
• identity-aware API gateways,
• and policy enforcement layers

start becoming strategically important.

Most people currently view MCP and A2A as integration standards.

I think that’s too narrow.

They are rapidly evolving into governance mechanisms for:

• trust,
• identity,
• context validation,
• data lineage,
• and policy enforcement between autonomous systems.

Because once agents begin communicating with other agents autonomously, governance can no longer remain static.

It must become dynamic and continuous.

3. Assurance – “How do we know the outcome is safe and trustworthy?”

This is the hardest problem of all. AI systems are probabilistic by nature. That means:

• bias,
• hallucination,
• overconfidence,
• drift,
• and unexpected behaviour

cannot be fully eliminated. Only managed.

This is why assurance becomes critical.

The AI Control Plane must increasingly support:

• behavioural validation,
• runtime verification,
• confidence scoring,
• policy alignment,
• output validation,
• adversarial testing,
• and continuous monitoring.

Security teams already understand this concept well.

We never assumed infrastructure was trustworthy forever. We continuously:

• monitor,
• validate,
• scan,
• test,
• and verify.

AI environments now require the same mindset. The difference is: we are no longer only securing infrastructure. We are securing autonomous decision-making.

That is a very different challenge

AI Is Creating a New Attack Surface

Now layer cyber security into the equation. Agentic AI environments introduce entirely new attack surfaces:

• prompt injection,
• model poisoning,
• malicious MCP servers,
• manipulated context windows,
• rogue plugins,
• compromised agents,
• and autonomous lateral movement.

Attackers are already experimenting with:

• adversarial prompts,
• deceptive reasoning chains,
• poisoned training repositories,
• and manipulated orchestration layers.

The scary part?

Many AI systems are deeply connected to:

• cloud platforms,
• SaaS environments,
• identity systems,
• APIs,
• and sensitive datasets.

A compromised AI agent with excessive permissions could become:

• an insider threat,
• a privileged automation engine,
• or an autonomous attack orchestrator.

Unlike humans, agents operate at machine speed. That changes dwell time dramatically.

The Market Is Currently Running AI Like a Formula 1 Car Without Brakes

Let’s be honest for a second.

Some organisations currently have:

• stronger governance for employee laptops than for autonomous AI systems connected to production data.

That should worry us.

The market is moving so quickly that many organisations are effectively:

• connecting copilots,
• exposing APIs,
• enabling agentic workflows,
• and granting broad access permissions

without fully understanding:

• trust boundaries,
• behavioural risks,
• operational blast radius,
• or governance implications.

It feels a bit like: “Let’s connect the AI to production and see what happens.”

Unfortunately, we are now starting to see what happens.

Where the Market Is Heading

Over the next few years, I believe we’ll see several major shifts.

1. AI Governance Moves Into Runtime

Governance will stop being static documentation and become:

• real-time,
• continuous,
• behavioural,
• and telemetry-driven.

2. AI Security and AI Operations Converge

The boundaries between:

• AI engineering,
• cyber security,
• governance,
• compliance,
• and platform operations

will blur rapidly.

3. AI Control Planes Become Standard Architecture

Just like CNAPP reshaped cloud security, AI Control Planes will emerge as:

• orchestration,
• governance,
• security,
• and assurance platforms for agentic ecosystems.

4. Regulators Will Move Faster

Expect increasing focus on:

• auditability,
• explainability,
• AI accountability,
• decision traceability,
• and operational governance.

Especially across:

• healthcare,
• financial services,
• government,
• and critical infrastructure.

5. Organisations Will Create AI Runtime Security Functions

This becomes a major growth area:

• AI runtime monitoring,
• agent governance,
• model assurance,
• behavioural analytics,
• and autonomous system security.

6. Organisation Will Focus on Managing Non-Human Identities

Many organisations are already struggling to govern human and machine identities at scale. Agentic AI introduces an entirely new category: autonomous non-human identities capable of making decisions, chaining actions and interacting with systems independently. The identity problem in AI may become larger than the model problem itself.

CyberArk shared a trajectory of 150:1 from about 50:1 today. That’s one human identity for every 150 non-human identities.

Secrets, tokens, certificates. Does not matter what your agents are using to authenticate and gain access, the pure trajectory over coming years will have to rethink access managment for the agentic era.

Final Reflection

AI is one of the most transformative technologies we’ve ever seen.

It has the potential to help humanity:

• predict faster,
• defend smarter,
• innovate quicker,
• and solve incredibly complex problems.

It allows us to become modern time travellers, anticipating the future before it fully unfolds.

But history repeatedly teaches us something important:

Every powerful technology eventually forces us to confront the consequences of acceleration without governance.

Agentic AI is no different.

The organisations that succeed over the next decade will not necessarily be the ones that deploy AI the fastest.

They’ll be the ones that:

• govern it properly,
• instrument it deeply,
• constrain it intelligently,
• and build sustainable guardrails around autonomy.

Because eventually, every organisation will face the same question:

“How much autonomous decision-making are we comfortable delegating to systems we still don’t fully understand?”

That conversation is no longer theoretical.

It has already started.

Key Takeaways and Learnings

• Agentic AI environments are volatile by nature and require new governance models.
• AI is enabling organisations to become “modern time travellers” through predictive intelligence and accelerated decision-making.
• The AI Control Plane is emerging as a foundational architectural layer.
• Three critical pillars will define successful AI governance: Auditability Governance Assurance
• MCP and A2A are evolving into governance and trust frameworks, not just integration standards.
• AI runtime security will become one of the fastest-growing cyber security domains between 2025–2027.
• Organisations must balance innovation speed with operational guardrails and sustainable governance.