September 2026
— Issue #6
Why cybersecurity must remain at the
centre of the AI debate
Patrick Houyoux LL.M.
President & Founder — PT SYDECO
There is so much excitement, anxiety and sometimes outright hysteria
surrounding Artificial Intelligence that we are beginning to lose sight of
something fundamental.
AI did not create a new digital universe.
It entered the one that already exists.
Every day, new discussions emerge about AI governance, AI ethics, AI
liability, AI-specific regulations, autonomous agents, responsible AI, trusted
AI, explainable AI and the laws that should govern all of them.
Many of these discussions are necessary.
But there is a danger in treating AI as if everything we have learned
during decades of cybersecurity suddenly became irrelevant.
It has not.
Before creating increasingly complex categories, regimes and regulatory
structures around AI, perhaps we should first return to a much simpler
question:
What exactly is an AI system when it operates inside an information
system?
The answer is less mysterious than the current debate sometimes suggests.
It is a digital actor.
AI DOES NOT OPERATE OUTSIDE THE
DIGITAL WORLD
An AI system uses computing infrastructure.
It accesses data.
It communicates through networks.
It may call APIs.
It may use credentials.
It may access files, databases, applications or external services.
An autonomous AI agent may even be authorised to execute commands, send
emails, move information, interact with customers, initiate transactions or
communicate with other systems.
In other words, AI operates through the same technological environment
that cybersecurity has been protecting for decades.
What has changed is not the existence of the digital environment.
What has changed is the nature of one of the actors operating inside it.
Traditional software executes predefined instructions.
An AI system may interpret information, make choices and, increasingly,
act without requiring a human instruction for every individual action.
That autonomy changes the risk.
But it does not remove AI from the domain of cybersecurity.
Quite the opposite.
It makes cybersecurity even more important.
THE REAL QUESTION IS NOT WHETHER AI
CAN THINK
Much of today's debate revolves around questions such as:
Can AI make decisions?
Can AI become autonomous?
Can AI behave unpredictably?
Who is responsible when an AI makes a wrong decision?
These are important questions.
But from a cybersecurity perspective, another question should come first:
What is the AI actually allowed to do?
Imagine an AI agent with access to:
- corporate
email;
- customer
databases;
- accounting
systems;
- confidential
documents;
- Internet
connectivity;
- API
credentials;
- cloud services;
- file-transfer
capabilities;
- system
administration functions.
If that AI performs an unauthorised action, we can spend months
discussing whether the model misunderstood an instruction, was manipulated by a
prompt injection, hallucinated, was badly trained or independently made the
wrong decision.
But there is an even more fundamental security question:
Why was the system architected in such a way that the AI was technically
capable of performing that unauthorised action in the first place?
That question belongs squarely to cybersecurity.
GOVERNANCE DEFINES WHAT AI SHOULD DO.
CYBERSECURITY DEFINES WHAT AI CAN DO.
This distinction is essential.
Governance establishes principles, responsibilities, policies and
acceptable behaviour.
Cybersecurity enforces boundaries.
Governance can say that an AI agent must not transfer confidential
corporate information outside the organisation.
Cybersecurity must ensure that the agent cannot do so.
Governance can say that an AI must not access financial records unless
specifically authorised.
Cybersecurity must enforce that access restriction.
Governance can establish who is responsible for an autonomous system.
Cybersecurity must ensure that every action performed by that system is
authenticated, authorised, logged and, when necessary, stopped.
Both disciplines are necessary.
But one should not replace the other.
A policy telling an AI what it should not do is not a security control.
A technical architecture preventing it from doing what it is not
authorised to do is.
WE ALREADY KNOW MANY OF THE PRINCIPLES
The cybersecurity community does not need to start again from zero simply
because the actor is now artificial.
We already know the fundamental principles:
Identity.
An AI agent should have an identifiable digital identity.
Authentication.
The system must know which agent, model, user or service is requesting an
action.
Authorisation.
An AI must only be able to perform actions explicitly permitted to it.
Least privilege.
An agent should receive the minimum permissions necessary to perform its
mission — nothing more.
Segmentation.
AI systems should not automatically have unrestricted access across an
organisation's infrastructure.
Zero Trust.
An AI should not be trusted merely because it operates inside the
corporate environment.
Every request should be evaluated according to identity, context, policy
and risk.
Monitoring.
AI actions must be observable.
Logging.
Important decisions and system actions must leave an auditable trail.
Containment.
An AI system whose behaviour becomes abnormal must be isolatable.
Data-loss prevention.
An AI should not be able to move sensitive information beyond authorised
boundaries.
Revocation.
Its credentials and privileges must be removable immediately.
And, ultimately:
If necessary, it must be possible to disconnect it.
None of these concepts was invented for Artificial Intelligence.
Yet all of them become even more important when software is capable of
deciding and acting.
AI DOES CREATE NEW RISKS
This does not mean that AI introduces nothing new.
It clearly does.
Prompt injection, manipulation of contextual information, poisoned
training or retrieval data, excessive agency, uncontrolled tool use, model
manipulation, non-deterministic behaviour and interactions between autonomous
agents create attack surfaces that did not exist in exactly the same form
before.
Cybersecurity must evolve accordingly.
But evolving cybersecurity is very different from pretending that AI
exists outside cybersecurity.
New attack techniques have appeared throughout the history of computing.
The answer has never been to abandon the fundamental principles of
security.
It has been to adapt them.
AI should be treated the same way.
AN AI AGENT SHOULD BE TREATED AS A
DIGITAL ACTOR
Perhaps one of the mistakes in the current debate is to treat autonomous
AI as something conceptually outside our existing systems.
I would propose a much simpler approach.
Treat an autonomous AI as a digital actor operating inside your
information system.
Give it an identity.
Define its role.
Limit its privileges.
Restrict its network access.
Control the data it can reach.
Monitor its behaviour.
Record its actions.
Prevent unauthorised communications.
Detect anomalies.
Contain it when necessary.
And never assume that intelligence creates trust.
Human employees are intelligent.
We do not therefore give every employee unrestricted administrator access
to the entire company network.
Why should an AI agent receive it?
RESPONSIBILITY REMAINS IMPORTANT — BUT
PREVENTION COMES FIRST
There will inevitably be legal debates about responsibility.
Who is liable when an autonomous AI causes damage?
The developer?
The model provider?
The company deploying it?
The integrator?
The operator?
The person who authorised its use?
These questions will require legal answers.
But cybersecurity cannot wait for all those answers.
A firewall does not need to determine legal liability before blocking an
unauthorised connection.
An access-control system does not need to understand moral responsibility
before refusing access to confidential data.
A segmented network does not need to determine whether an AI had
malicious intent.
It simply prevents the system from reaching what it was never authorised
to reach.
This distinction is fundamental.
Law determines responsibility after an event.
Cybersecurity must reduce the possibility of the event occurring in the
first place.
LET US PUT CYBERSECURITY BACK AT THE
CENTRE
AI deserves serious governance.
It deserves serious ethical reflection.
It may require adjustments to existing laws and, in some circumstances,
new rules.
But we should resist the temptation to treat Artificial Intelligence as
if it had created an entirely separate technological universe requiring us to
reinvent every principle from the beginning.
It has not.
AI is software.
It operates through infrastructure.
It uses permissions.
It accesses information.
It communicates through networks.
And increasingly, it acts.
That makes AI not less relevant to cybersecurity, but more relevant.
The question should therefore not only be:
How do we govern AI?
We must also ask:
How do we control what AI is technically capable of doing?
Because however sophisticated Artificial Intelligence becomes, one
fundamental principle remains:
AI IS NOT A NEW CYBER UNIVERSE.
IT IS A NEW ACTOR INSIDE THE EXISTING ONE.
And like every other actor inside that universe, it must operate within
defined boundaries.
DO NOT TRUST IT
BECAUSE IT IS INTELLIGENT.
DO NOT FEAR IT
BECAUSE IT IS AUTONOMOUS.
SECURE IT
BECAUSE IT IS CONNECTED.
The Sydeco Security Brief
Cybersecurity intelligence, architecture and doctrine from PT SYDECO.
