Pakistan is spending heavily on digital banking, e-health, online government services and, more and more, on artificial intelligence. The public conversation is mostly about productivity and growth. Whether our institutions can actually secure what they are building gets far less airtime, even though every promised benefit depends on it.
AI has changed the economics of cybersecurity, for attackers and defenders both. Vulnerability research that used to take a specialist weeks can now be done in days. Code is being shipped faster than anyone can review it. AI agents are being plugged into email, databases and cloud accounts, usually with more access than they need.
None of this arrives on a clean slate in Pakistan. Security practice in most of our institutions never caught up with digitisation in the first place. We still lean on periodic audits, compliance paperwork and the purchase of security products. They have their uses. What they cannot do is stand in for continuous technical oversight, a working incident response capability, and experienced people who have the authority to make calls.
We have already seen what happens when public systems are disrupted. In June 2026, Islamabad’s Capital Development Authority reported a cyberattack on its property and water billing system, interrupting online services. The technical cause has not been established publicly, but the consequences were plain enough.
Consider PHP. In 2026, researchers from Calif, Wiz and OpenAI found serious vulnerabilities in it, some allowing remote code execution. PHP is hardly obscure software; a large part of the web runs on it, and developers and researchers have been combing through its code for decades. The PHP2Shell flaws were there anyway.
Anthropic’s work with Mozilla points in the same direction. Using Claude, researchers turned up 22 vulnerabilities in Firefox in two weeks, 14 of them rated high severity. Could skilled humans have found them? In principle, yes. But looking has become cheap, and with AI help it now makes sense to go back through mature codebases, test strange edge cases and follow leads nobody used to have the patience for.
That should worry anyone running older systems in a bank, hospital or government department. Years in production prove very little about security, and audits miss things. In the eMudhra certificate case, weak cryptographic keys went undetected for about 18 months, through established procedures and independent audit cycles alike. An audit reflects what could be checked on the day it was done. Systems keep changing after that, and so do the techniques available to whoever is probing them.
Conventional security tools are fairly good at catching known weaknesses. Where AI agents help is with the slower work of watching how a system responds, spotting odd behaviour and figuring out how several small problems might be chained into one large one. Serious breaches are rarely the product of a single brilliant exploit. Usually it is a run of ordinary failures, each unremarkable on its own.
Binance is a great example. On 29 October 2026, the exchange reported a security incident involving roughly $6.34 billion in digital assets. Its preliminary investigation says attackers reached engineers’ machines through malware and social engineering, a fake hiring process among them, and then used stolen credentials, compromised internal AI tooling and access to wallet infrastructure to push through unauthorised transactions.
There is another detail worth understanding. On Ethereum, the attackers used EIP-7702 authorisations to delegate control to a malicious smart contract. Those permissions remained active even after normal wallet services were suspended, so stopping the attack required more than shutting down the affected system.
The cryptography was not broken, and quantum computing had nothing to do with it. The attackers simply went after the people and systems trusted to run the infrastructure, and once they controlled the machines that approve transactions, the strength of the encryption was beside the point. Plenty of organisations outside crypto run on the same kind of trust.
The software is also going live faster. A small group of youngsters can now build what once took a large team, which is welcome news for Pakistan’s tech sector if security review keeps up. So far it hasn’t. Writing code is getting cheaper much faster than it is being understood or governed.
In many cases an AI-generated application can do exactly what it was asked to do and still have unreliable authorisation. An API might return the right data while letting one customer see another’s records; a cloud integration might run without a hitch on privileges far wider than it needs. I am not saying that human developers won’t make these mistakes. The change lies in the volume, since a lot more software now reaches production without quality vetting. Annual penetration testing retains its place, but applications that change every few weeks, and keep picking up new APIs and third-party integrations along the way, need testing throughout development.
In my experience things get riskier when AI stops helping people write software and starts acting on their behalf. A tool that summarises documents can only do so much harm. An agent able to read customer records, alter cloud settings or send data outside the organisation can do unquantifiable damage. A malicious instruction seeded into an email or document may be enough to steer a poorly protected agent, which is why permissions, access controls, logging and human approval for sensitive actions deserve far more scrutiny.
What we must understand is that these risks cut across Pakistan’s critical infrastructure, though sectors vary in how much damage they would take and how well they could recover. Banking and telecoms are generally better resourced and more closely regulated. They still depend on software vendors, cloud providers, staff laptops and third-party integrations, and an attacker needs only one weak entry point.
The recent attack on French bank Crédit Mutuel Arkéa makes the third-party risk harder to dismiss. The bank reported around €100 million in fraudulent transfers and said the attackers reached it through a service provider’s platform. How that access was obtained is still under investigation, but an organisation’s security increasingly depends on people and systems it does not directly control.
Pakistan’s banking sector has at least begun addressing resilience more systematically. The State Bank launched its Cyber Shield strategy this year, and 34 financial institutions took part in an industry-wide cyber drill. That is useful progress. The same attention is harder to find across healthcare, government services and the suppliers they depend on.
Healthcare concerns me most. A breach there hits privacy and patient care at the same time. Hospitals hold diagnoses, prescriptions and entire medical histories, and while a stolen password can be reset, a stolen medical history cannot be taken back. In the United States, the Change Healthcare attack affected information on about 192.7 million people. In Britain, the Synnovis ransomware attack disrupted pathology services and delayed thousands of appointments and procedures. Pakistan’s health systems are becoming more connected by the year.
We have had warnings at home too. Maroof International Hospital in Islamabad suffered a ransomware attack in 2020 that disrupted its computer systems. Linking more medical records without improving security and recovery arrangements would only increase what is at stake.
The One Patient, One ID initiative could be genuinely useful, giving authorised providers a patient’s history across institutions. Linking records raises the stakes, though. Without proper safeguards, a weakness in one integration could expose patient data held by many providers, and any AI tool that processes documents, summarises records or supports staff has to sit inside the same controls as the systems it reaches into.
With this much sensitive data being collected and shared, it is hard to explain why Pakistan still has no comprehensive personal data protection law in force. Healthcare needs clear cybersecurity requirements, independent assessments, mandatory incident reporting, and someone answerable when patient data leaks. National CERT has a coordinating role, but it cannot secure every hospital, laboratory and vendor in the country. Health authorities and sector regulators will have to carry that responsibility themselves.
Energy and industry are a different case, because keeping operations running comes first. Much of their operational technology sits on ageing equipment that is hard to patch and cannot be tested aggressively. Remote monitoring, cloud analytics and AI-driven operations are pulling these environments closer to ordinary IT networks, and a breach there can have physical consequences. Protecting them takes people who understand cybersecurity and industrial control systems together.
Across all of these sectors the constraint is the human factor. AI can generate findings faster than organisations can triage them and write code faster than teams can review it. It can propose attack paths, but an experienced researcher still has to verify them. Is a vulnerability actually exploitable? Will an emergency patch knock out an essential service? Do three minor issues add up to one serious one? Answering those questions takes experience, and institutions that are prepared to act on what experienced people tell them.
Pakistan does have capable researchers, engineers and security professionals, some of them working on complex international projects. Developing and keeping them is harder. A good exploit developer, incident responder, malware analyst or industrial security engineer takes years to produce, and training more entry-level staff, necessary as it is, does nothing for the shortage at the senior end.
Even where the expertise exists, it often goes unused. Highly technical posts are too often filled as if they were general administrative jobs. An administrator may be excellent at governance, coordination and policy and still be the wrong person to assess a security architecture or investigate an advanced threat. Short postings compound the problem: by the time an officer knows enough to be effective, they have been transferred, and the institution starts over.
To be absolutely clear, I am not suggesting engineers should replace administrators or policymakers. Good governance needs both, with responsibilities clearly divided. Seniority and technical competence are simply different things, and specialists need a real say in decisions on architecture, procurement and regulation. Dedicated technical career tracks, secondments from industry and independent advisory positions would all help. Retaining that expertise, and giving it weight, belongs on the security agenda.
Digitisation should go ahead. Security practice has to keep up with it, and a handful of practical steps would make a real difference.
Critical organisations need a clear inventory of what they run, from systems and software dependencies to cloud services and the suppliers holding privileged access, along with a tested process for urgent patching when a serious flaw turns up. Testing should be tied to change, so that major updates, new APIs and new third-party integrations get checked as they happen and not at the next annual compliance cycle. AI agents handling sensitive data need tight permissions and proper monitoring, with separate approval for anything high-risk.
Healthcare cannot wait; a data protection law and clear sector-level responsibility for cybersecurity are long overdue. Security also belongs in procurement, with government assessing suppliers, maintenance arrangements and third-party access before a contract is signed. And technical expertise needs more weight than it currently gets, which means investing in experienced specialists, holding on to them and bringing them into decisions on architecture, procurement and security.
Organisations can buy expensive security products, commission audits and collect certifications. If serious vulnerabilities still go unfixed because nobody understands the risk or has the authority to act, that money has bought very little.
AI will help cyber defenders find vulnerabilities, investigate incidents and automate routine work. It will help attackers in much the same ways. Vulnerabilities are surfacing faster, software changes more often and systems are more tightly interconnected than they have ever been, and the real question for Pakistan is whether its institutions can respond in time. The talent exists. It needs to be retained, brought closer to the decisions that matter and given the authority to act, preferably before a major incident leaves no choice.
