Kenneth Hawrylak, November 15 2026

AI as Attacker, Defender and Target

The third front in cybersecurity that too many people overlook.

Blue and red cybernetic armies clash above separate AI computer systems as green digital corruption spreads through their infrastructure and compromises their soldiers.

Artificial intelligence is changing cybersecurity. That much is clear.

Attackers are using it to improve their techniques. Defenders are using it to identify threats and respond more efficiently. Every day, we hear about the growing technological arms race between the two.

But there's a third part of this equation that doesn't get nearly enough attention.

What happens when the AI systems themselves become the target?

And perhaps more importantly, what happens when those systems are given enough access and authority to cause damage without an attacker being involved at all?

That's where the conversation gets interesting.

AI as Attacker: The Offensive Advantage

Cybercriminals don't need to invent entirely new attacks to benefit from artificial intelligence.

Many existing techniques can become faster, cheaper, or more convincing with AI assistance.

Consider phishing.

Traditional phishing emails often contain obvious spelling mistakes, awkward wording, or generic messages sent to thousands of recipients.

AI can help attackers craft convincing messages tailored to a particular person, company, or situation. Combine that with publicly available information, and a fraudulent email can become considerably harder to recognize.

AI can also assist attackers with reconnaissance, vulnerability research, malicious code development, and impersonation through synthetic audio or video.

But there's an important distinction.

AI doesn't automatically make an attack successful.

An AI-generated phishing email still needs someone to believe it. An identified vulnerability still needs to be exploitable. A sophisticated-looking attack isn't necessarily a technically effective one.

AI can amplify an attacker's capabilities, but the underlying weaknesses being exploited are often very familiar.

AI as Defender: Strengthening Security

Fortunately, defenders have access to many of the same technological advantages.

AI-assisted security tools can help analyze enormous quantities of information, recognize unusual activity, and bring suspicious events to an analyst's attention.

For example, an organization might use AI-assisted monitoring to identify unusual account activity, investigate potentially malicious emails, or correlate security events across multiple systems.

Some security platforms can also automate responses, such as isolating a compromised computer or temporarily blocking suspicious activity.

These capabilities can be valuable, particularly when security teams face more alerts than they can reasonably investigate manually.

But we need to be careful about what we assume those systems can do.

Detecting something unusual is not the same as proving something malicious has happened.

An employee signing in from an unfamiliar location might indicate a stolen account. It might also mean they're travelling.

A legitimate application might behave differently after an update.

An AI system might correctly identify an unusual pattern but incorrectly explain its cause.

And if that system is authorized to take action automatically, an incorrect conclusion can quickly become an operational problem.

AI can assist security decisions. That doesn't make its conclusions infallible.

AI as Target: The Missing Third Front

Now consider what happens when we start connecting AI systems to the applications and information businesses rely on.

An AI assistant might be allowed to read emails, search company documents, access customer records, update databases, or interact with other software.

Some systems, commonly called AI agents, can go further. They can select tools, carry out sequences of tasks, and perform actions on behalf of their users.

That makes them useful.

It also creates new opportunities for things to go wrong.

An AI system can be manipulated through malicious instructions hidden in information it processes. Its data can be exposed through poorly controlled access. Compromised components or poisoned information can influence what it produces.

And an AI application with excessive permissions may be capable of actions that should never have been delegated to it.

This isn't about AI suddenly developing malicious intentions.

It's about trusted software processing untrusted information while possessing access to systems that matter.

The security industry already recognizes these risks. OWASP, an organization known for its application-security guidance, identifies issues such as prompt injection, sensitive information disclosure, excessive agency, and compromised AI dependencies.

The technology may be new, but many of the underlying security principles are not.

When Your Own Systems Turn Against You

Imagine a small business using an AI assistant to help process supplier invoices.

The assistant reads incoming invoices, checks amounts against purchase orders, and prepares information for the accounting system.

So far, nothing particularly alarming.

Now imagine an attacker sends a convincing invoice containing hidden instructions intended for the AI assistant.

The instructions attempt to make the assistant replace the supplier's banking information with an account controlled by the attacker.

This technique is called prompt injection.

The attacker isn't issuing an authorized command to the business. They're placing instructions inside a document that the AI is supposed to treat as information.

If the application fails to maintain that distinction, and the assistant has permission to modify supplier records, the consequences could be serious.

The assistant might attempt to change the payment information using access legitimately assigned to it.

Whether the attempt succeeds depends on the application's security controls.

A properly designed system should prevent an invoice from authorizing changes to banking details. Sensitive changes should require independent verification and appropriate approval.

But consider the underlying problem.

The attacker didn't necessarily need to compromise the accounting application directly.

They tried to manipulate something that already had legitimate access.

The vulnerability is the failure to keep untrusted information separate from trusted authority.

That's a security problem worth understanding whether your business uses AI or not.

The Battlefield Beneath the Battlefield

Our accompanying illustration depicts two opposing armies, one blue and one red.

Beneath them are two separate AI systems, connected to their respective forces.

But something else is spreading through the infrastructure: green digital corruption, working its way through connections and into the soldiers themselves.

Some soldiers have compromised limbs turning against their own side.

It's deliberately surreal, but the underlying idea is practical.

The red and blue forces represent attackers and defenders using AI.

The underground systems represent the technology, data, permissions, and infrastructure they depend on.

The green corruption represents vulnerabilities and failures that can undermine either side.

Sometimes that means deliberate manipulation by an attacker.

Sometimes it means a misconfiguration, excessive permissions, unreliable information, or poorly controlled automation.

These are different causes, and they shouldn't be confused. But they can produce a similar outcome: a trusted capability being used incorrectly or against its owner's interests.

And there's another complication.

Many organizations don't operate their AI systems entirely themselves.

They depend on external providers, cloud services, software integrations, and third-party models.

That introduces additional security and operational dependencies.

What happens if a provider experiences an outage? An integration is compromised? A service changes its permissions or behaviour?

Understanding those dependencies is part of understanding your security exposure.

Securing the Third Front

The good news is that protecting AI-enabled systems doesn't require abandoning everything we've learned about cybersecurity.

Quite the opposite.

Many established security principles become even more important when software can make decisions and take actions on our behalf.

1. Know where AI is being used.

Identify which AI tools employees use, what information those tools process, and which business systems they're connected to.

You can't properly secure something you don't know is there.

2. Apply least privilege.

Give AI applications access only to the information and functions required for their assigned tasks.

An assistant that summarizes invoices probably doesn't need permission to change supplier banking information.

3. Separate information from instructions.

Emails, webpages, documents, and other external material should be treated as potentially untrusted.

Information contained in those sources must not automatically become authorization to perform an action.

4. Validate results and control consequential actions.

AI-generated conclusions should be checked against reliable evidence when the stakes justify it.

Financial changes, account permissions, security configurations, and other significant decisions should have safeguards that operate independently of the AI's own judgment.

5. Monitor activity and prepare for failure.

Maintain useful records of what AI-enabled systems access and do.

Know how to revoke their permissions, interrupt automated processes, investigate unexpected actions, and recover from mistakes.

A system that can't be stopped, examined, or corrected is a system with unnecessary risk.

The Real Question Isn't Whether AI Is Good or Bad

AI is neither a magic cybersecurity solution nor an unstoppable digital enemy.

It's technology.

Technology that can help people work more efficiently, detect threats, and solve difficult problems.

Technology that can also be misused, compromised, misconfigured, or trusted beyond its actual capabilities.

As AI becomes more deeply integrated into business operations, the consequences of getting those things wrong can grow.

The challenge isn't simply defending against attackers who use AI.

It's also ensuring that the AI systems we choose to trust are appropriately secured, monitored, and controlled.

We've spent decades developing security principles around identity, access, authorization, verification, and accountability.

There's no good reason to abandon those principles just because the software now appears capable of reasoning.

AI as Attacker. AI as Defender. AI as Target.

Three perspectives on the same evolving cybersecurity landscape.

And one question every organization adopting AI should be asking:

We're giving AI systems increasingly powerful capabilities. Are we giving equal attention to securing them?

Further reading

OWASP — Top 10 for Agentic Applications

Canadian Centre for Cyber Security — Careful Adoption of Agentic AI

NIST — Generative AI Risk Management Profile


Digital Drive IT Solutions

Demystifying technology so people can make better decisions.

Written by

Kenneth Hawrylak

Older The 3-2-1 Backup Rule: Why One Backup Isn't Enough