For years, companies treated artificial intelligence as a productivity race.
Who could deploy it faster?
Who could automate more work?
Who could build the better chatbot?
Who could use AI to cut costs and increase revenue?
That conversation is changing.
Now there is another race happening underneath it:
Who can defend their company against AI-powered attacks before someone else uses AI against them?
That was the warning from Citigroup Chair and CEO Jane Fraser at the 2026 Qatar Economic Forum, where she said companies are racing to strengthen their defenses around artificial-intelligence models. The Qatar Economic Forum lists Fraser among its featured 2026 speakers and has made “Technology Hype & Reality” one of its central themes.
Her warning arrives at an important moment.
AI is no longer simply generating emails, writing code and summarising documents.
The newest systems can increasingly reason through complicated problems, use tools, interact with external systems and conduct multi-step operations.
That creates an uncomfortable possibility:
The same technology being deployed to protect companies can potentially be used to attack them.
The hacker is getting an AI assistant
Traditional cyberattacks often require expertise.
An attacker has to discover vulnerabilities, understand the target, develop an exploit, maintain access and move through a network.
AI can compress parts of that process.
A September 2026 paper from the Bank for International Settlements' Financial Stability Institute warned that frontier AI models can increasingly identify vulnerabilities, develop exploits and conduct complex multi-step cyber operations autonomously. The paper says this can reduce the expertise, time and resources required for sophisticated attacks.
That changes the economics of cybercrime.
A highly skilled hacker has traditionally been a scarce resource.
If AI can automate significant portions of the work, the limiting factor could increasingly become access rather than expertise.
And financial institutions have an especially large target on their backs.
Banks hold money.
They hold sensitive customer information.
They operate enormous digital infrastructures.
They connect to payment networks, exchanges, governments, businesses and millions of customers.
A successful attack can therefore have consequences far beyond one compromised computer.
The banks cannot simply switch AI off
This is the paradox.
Financial institutions increasingly need AI because their competitors are using it.
AI can analyse enormous quantities of information, detect suspicious transactions, automate customer-service operations, assist employees and help identify fraud.
The World Economic Forum reported in June that financial institutions are moving from AI experiments toward scaled deployment, while simultaneously putting greater emphasis on trust, governance and human oversight.
The Financial Stability Board has similarly warned that rapid AI adoption can introduce or amplify risks in financial institutions and has proposed 12 governance practices covering the AI lifecycle.
So banks face a difficult calculation.
Move too slowly and they may lose the productivity and competitive advantages of AI.
Move too quickly without sufficient controls and they could introduce new security, operational and regulatory vulnerabilities.
The solution cannot simply be:
“Don't use AI.”
It increasingly has to be:
“Use AI, but build an equally serious system around controlling it.”
The defensive AI race
There is already evidence that the technology industry sees this as an emerging emergency.
In August, more than 100 companies, including OpenAI, Anthropic, Microsoft, Alphabet, Amazon, Capital One, Cloudflare, CrowdStrike, IBM, Mastercard, Oracle, Robinhood, Shopify and Visa, called for a defensive surge against AI-enabled cyberattacks.
The companies warned that increasingly capable AI systems could make AI-enabled cyberattacks substantially more widespread and urged organisations to make cyber defence an immediate leadership priority.
That is an extraordinary development.
Some of the world's largest technology companies are effectively acknowledging that the capabilities they are helping create could also increase the capabilities available to attackers.
The cybersecurity industry therefore faces a strange new competition:
AI versus AI.
An attacking model searches for the weakness.
A defensive model searches for the attack.
One attempts to exploit a vulnerability.
The other attempts to detect it before exploitation.
One searches for the shortest route into a network.
The other attempts to close every possible route.
And both sides can potentially operate at machine speed.
The old cybersecurity model is becoming insufficient
Traditional cybersecurity has often been built around known threats.
A company identifies a vulnerability.
It patches it.
It installs security software.
It monitors suspicious activity.
It responds when something happens.
But AI changes the speed of the game.
An attacker using an advanced model may be able to discover and exploit weaknesses faster than human security teams can manually investigate them.
The BIS has specifically warned about the compression of the time between vulnerability discovery and exploitation. It identifies unpatched software as a major initial-access pathway and argues that AI can make exploit chaining more efficient.
That means companies may have less time to react.
Security increasingly has to become predictive and automated.
Instead of waiting for someone to attack, systems need to continuously look for weaknesses.
Instead of analysing suspicious behaviour hours later, defensive AI needs to detect it almost immediately.
Instead of relying entirely on a security team watching dashboards, companies need machines capable of monitoring enormous numbers of events simultaneously.
But putting AI inside the defence system creates another problem
Who watches the AI?
This may become one of the most important questions in corporate cybersecurity.
An AI security system could receive enormous privileges.
It might monitor corporate networks.
It could access logs.
It might isolate machines.
It could block connections.
Eventually, highly autonomous security agents could potentially make decisions about threats without waiting for a human.
That creates a new attack surface.
If an attacker compromises the defensive AI itself, the consequences could be enormous.
And if the defensive model makes a mistake, it could potentially shut down legitimate business activity.
Imagine an AI security system deciding that a major payment system looks suspicious.
The machine blocks it.
The bank discovers an hour later that the activity was legitimate.
The system was technically functioning exactly as designed.
But the business was still disrupted.
This is why AI security cannot simply mean installing a smarter model.
It requires governance around the model.
Control is becoming the new AI advantage
A recent World Economic Forum analysis described this emerging problem as a question of control.
Companies are increasingly asking whether they can scale AI while retaining control over their technology, intelligence and economics.
The WEF specifically highlighted the need for systems that can govern, orchestrate, secure, monitor and manage growing “digital workforces” of AI agents, with observability and audit trails built into the infrastructure.
That concept could become one of the defining corporate technologies of the next few years.
Not just AI.
AI control systems.
Companies will need to know:
What is the model doing?
What data can it access?
What decisions can it make?
Which systems can it change?
Who authorised the action?
Can the action be reversed?
What happens if the model behaves unexpectedly?
And perhaps most importantly:
Can the company shut the system down?
The financial sector has even more at stake
For a bank, AI security is not simply an IT problem.
It is a financial-stability issue.
The BIS argues that frontier AI could increase cyber risks for financial institutions by shortening remediation windows, increasing the probability of breaches and amplifying dependence on third-party technology providers.
That last point is particularly important.
A bank may not build its own AI model.
It may rely on an external model provider.
It may use a cloud provider.
It may connect several AI agents to internal systems.
It may use third-party cybersecurity products.
Suddenly, one institution's digital security depends on a chain of other companies.
That creates a new version of systemic risk.
The weakest link may not be inside the bank.
It may be somewhere in the technology ecosystem surrounding it.
The AI security paradox
There is something almost circular about this new era.
Companies are building AI because it can make them more efficient.
Then they discover that AI creates new vulnerabilities.
So they build AI security systems.
Those security systems become increasingly autonomous.
That creates new risks.
So companies build systems to monitor the AI monitoring the systems.
And eventually the corporate technology stack begins to resemble a collection of machines watching other machines.
The human role changes.
Humans increasingly become the people setting permissions, defining objectives, reviewing exceptions and deciding where machines are allowed to act.
That does not make humans irrelevant.
It makes human governance more important.
This is not just a Silicon Valley problem
For African companies, including Nigerian banks and fintechs, the lesson is particularly relevant.
Africa's financial sector is rapidly digitising.
Mobile banking, fintech platforms, digital payments and online lending have expanded the amount of financial activity taking place through software.
That creates enormous opportunities for AI.
But it also creates an enormous attack surface.
A Nigerian bank deploying AI cannot think only about how much money the technology can save.
It also needs to ask what happens if the model is manipulated, compromised, misconfigured or granted excessive access.
The same applies to fintech companies.
AI can help detect fraud.
But attackers can also use AI to generate more convincing scams.
AI can help write secure code.
But AI can also help identify vulnerabilities.
AI can improve customer service.
But it can also create convincing automated impersonation.
The technology is neutral in capability.
The security architecture determines who benefits from that capability.
The next corporate AI race may not be about the smartest model
This is perhaps the biggest takeaway from Fraser's warning.
The future may not belong simply to companies with the most powerful AI.
It may increasingly belong to companies that can deploy powerful AI without losing control of it.
That means security teams will become more important, not less.
Boards will need to understand AI risk.
Chief information security officers will have to understand AI agents.
AI teams will have to understand cybersecurity.
Regulators will increasingly have to understand how autonomous systems interact with financial institutions.
And companies will need detailed records of what their AI systems are doing.
Because once an AI agent can act, rather than merely recommend, accountability becomes much more complicated.
The arms race has already begun
The most important part of Jane Fraser's warning is not that companies are worried about AI.
They have been worried about AI for years.
It is that defence itself is becoming an AI race.
The attackers are getting faster.
The defenders are getting faster.
The models are becoming more capable.
And the distance between an AI system that can assist a human and one that can independently perform complicated operations is narrowing.
OpenAI has said it expects increasingly capable models to eventually perform much of the work involved in cybersecurity, while emphasising monitoring, alignment and security controls as three complementary safeguards.
That may be where the industry is heading.
Machines defending machines.
AI hunting vulnerabilities.
AI detecting AI-generated attacks.
AI monitoring AI agents.
And humans sitting above the system, trying to ensure that the machines remain under control.
The question for corporate leaders is therefore changing.
It is no longer simply:
“How quickly can we adopt AI?”
It is:
“How quickly can we build the security, governance and control systems required to survive an AI-powered world?”
Because the AI revolution is no longer approaching the cybersecurity battlefield.
It has entered it.
Comments
Post a Comment