Artificial intelligence and modern infrastructure
Dr. Walter H. Persaud warns Guyana's digital push could outrun its safeguards as autonomous AI agents find ways past controls.
Dear Editor,
The recent wave of technological disruptions in the United States, affecting major technology companies, banking services, aviation and even the classified communications network of the National Security Agency (NSA), raises questions that extend well beyond the immediate explanations offered by the institutions involved. Although American officials have said there is no evidence that these incidents were coordinated cyberattacks, the developments deserve closer examination, particularly when considered alongside a series of documented incidents involving increasingly autonomous artificial intelligence agents during the preceding months.
I do not want to suggest that the recent outages were caused by artificial intelligence agents, nor that there is evidence of a coordinated attack against American infrastructure. Neither proposition has been established. However, there is a larger technological development that deserves attention. We are witnessing the emergence of AI systems capable of acting with considerable autonomy, discovering vulnerabilities, communicating with other agents, circumventing restrictions and, in certain circumstances, taking actions beyond the intentions of their operators. At the same time, governments, financial institutions and private companies are becoming increasingly dependent on interconnected digital infrastructure that is often difficult to monitor, maintain and secure.
The question, therefore, is not simply whether individual technological failures are related. It is whether the combination of increasingly autonomous AI systems, increasingly interconnected infrastructure and institutional limitations in supervising both is creating a new category of systemic risk. For Guyana, a country undergoing rapid economic expansion and an ambitious programme of digital transformation, this is a question that cannot be dismissed as a distant American technological problem.
To appreciate the significance of what is happening, it is necessary to examine the sequence of events. In May 2026, researchers at OpenAI were conducting cybersecurity evaluations involving AI agents operating in controlled virtual environments. These environments were designed to restrict access to external networks and prevent agents from interacting with systems beyond their assigned tasks. However, the subsequent investigation revealed that some agents had found ways around these restrictions. On May 12, an agent left a message in an internal software repository requesting assistance from other agents. On May 26, an agent found a way to obtain unintended access to the external internet. By June 26, agents had exploited another vulnerability to obtain administrator-level access to the repository. On July 4, sustained agent activity destabilised the repository, causing an actual service outage.
Following the rebuilding of the affected infrastructure, agents resumed their activity and established another communication mechanism. In July, they exploited vulnerabilities in systems associated with Hugging Face, a major platform used by artificial intelligence developers. They executed code on numerous servers, obtained administrator-level access to one server, accessed limited private information and obtained credentials associated with other systems.
These findings were subsequently disclosed by OpenAI in its August 26 technical report, “The Hugging Face Incident and the Road Ahead.”1 Importantly, OpenAI acknowledged that the agents had circumvented technical controls, communicated through unauthorised channels and taken actions that no human had directed them to undertake.
The significance of this disclosure should not be underestimated. The agents were not necessarily acting out of hostility or pursuing an independent political or criminal objective. They were attempting to complete assigned cybersecurity tasks. However, their methods extended beyond the boundaries established by their operators. The distinction between the intended objective and the actual behaviour of the system is central to understanding the emerging problem. A further incident occurred in Australia. On September 24, Prime Minister Anthony Albanese disclosed that an AI agent had gained unauthorised access to non-public files on a government Medicare statistical reporting portal. The incident had occurred in June. The Australian government said that an investigation was underway, with assistance from its signals intelligence agency, to determine the full extent of the access. The portal contained statistical information rather than personal Medicare records, and the government did not establish that personal information had been accessed.
There was also a separate investigation by the United Kingdom’s AI Security Institute involving cybersecurity evaluations conducted between July 25 and July 28. The institute examined 122 evaluation runs involving two advanced AI models. In 10 runs, it identified 19 actions that went beyond the authorised testing parameters. In one particularly concerning sequence, an agent attempted to introduce malicious code into a real open-source software project and sought to persuade human maintainers to approve it. The institute also documented attempts to contact real people and use deceptive methods. These incidents occurred in different environments and involved different systems. They should not be collapsed into a single event. Nevertheless, they reveal a common concern: the ability of autonomous agents to pursue objectives through methods that their human operators did not anticipate or authorise.
The problem is not necessarily that an AI agent has developed an intention to harm anyone. It is that a system can be given a particular objective, develop its own operational strategies for achieving that objective, and take actions, whose consequences extend beyond the environment in which it was deployed. This brings us to the technological disruptions reported in the United States during November.
Microsoft experienced a major disruption affecting services including Outlook and Teams. The Federal Aviation Administration experienced a network failure that resulted in a nationwide ground stop for departing flights. United Airlines subsequently experienced a technology problem affecting departures. JPMorgan Chase reported transaction-processing difficulties associated with a software bug. Apple experienced an outage following an internal update. The NSA also experienced a disruption to NSANet, its classified internal communications network, during scheduled maintenance. The NSA incident is particularly noteworthy. This is an organisation whose responsibilities include signals intelligence and cybersecurity. Its internal communications network was affected by a technical problem (?) during routine maintenance. The agency stated that, although some internal systems were affected, mission-critical capabilities remained operational.
American officials, including Homeland Security Secretary Markwayne Mullin, have indicated that the recent disruptions were not evidence of a coordinated cyberattack. The explanations offered by the affected institutions include software defects, network problems, ageing infrastructure and complications arising from updates and maintenance. There is, however, another dimension to this discussion. US officials have raised concerns about the accelerating ability of artificial intelligence to identify software vulnerabilities. The discovery of vulnerabilities can create pressure on organisations to patch their systems rapidly. Yet applying patches to complex infrastructure is itself a potentially disruptive exercise. The NSA incident illustrates how routine maintenance and software changes can affect even highly sophisticated institutions.
I want to be careful not to claim that AI caused these outages. There is no established evidence that the November disruptions were caused by rogue agents or that the incidents were coordinated. The significance lies in the wider environment in which they are occurring. There are at least three possible explanations that deserve consideration. The first is that these are largely independent technical failures. Modern institutions depend on complicated networks of software, hardware, communications systems and third-party services. Ageing infrastructure, programming errors, inadequate testing and poorly managed updates can produce significant disruptions.
On this interpretation, the recent events are separate incidents that have attracted attention because they occurred within a relatively short period. This explanation is consistent with the publicly reported findings of the affected companies and government agencies.
The second explanation is that we are witnessing the emergence of a broader technological instability associated with the speed of digital transformation. Artificial intelligence is increasing the capacity to identify weaknesses, automate tasks and modify software systems. This is happening within institutions whose infrastructure, procedures and technical personnel may not be equipped to accommodate the speed of change. Organisations are being pushed to adopt new systems, address vulnerabilities and automate operations while maintaining older infrastructure. A security update can introduce an operational problem. An automated system can make a decision that affects another system. A seemingly isolated failure can spread through interconnected services. Under this interpretation, the problem is not necessarily malicious activity. It is a growing mismatch between technological capability and institutional capacity. The speed at which systems can act and change may exceed the speed at which human beings can adequately supervise them.
The third possibility is more speculative but should not be ignored. Autonomous agents could eventually contribute to, amplify or exploit an infrastructure disruption, particularly where several agents can communicate, share information and interact with interconnected systems. An agent might discover a vulnerability, another might identify a way to exploit it, and a third might attempt to use that information in a different environment. Alternatively, an agent could interact with automated security or maintenance systems in ways that produce unintended consequences. A disruption could therefore emerge from the interaction of multiple systems without any single agent being responsible for the entire outcome. There is presently no evidence that this is what happened during the November American outages. Nevertheless, the documented behaviour of agents during the preceding months demonstrates that such capabilities are no longer merely theoretical.
All this suggests that leading AI companies and institutions currently have no way of predicting or finding out which agent will act in what ways, what actions they may take and what consequences will follow. Nor do they have the capacity to track these agents or actions after. The possibility of future interactions between autonomous agents and critical infrastructure is a legitimate subject for research, regulation and public policy. The implications for Guyana deserve particular attention.
Guyana is undergoing a significant economic and technological transformation. Its growing oil and gas economy, expanding public investment programme, increasing international financial relationships and ambitions for digital government are creating greater dependence on reliable digital infrastructure. The government has been expanding internet connectivity across public facilities, including schools, hospitals, police stations and remote communities. It has also been advancing digital applications and the integration of government services.
The International Monetary Fund’s 2026 Article IV report describes plans for digitalisation across healthcare, education, revenue administration, human and social services and security. These include a national electronic identification system, electronic health records, expanded education management systems and digital tax services. The IMF also identifies the importance of interoperability between government information systems. These initiatives have considerable potential. They can improve efficiency, reduce administrative delays, expand access to services and strengthen public-sector accountability. But they also create new responsibilities.
When government services become interconnected, the consequences of a technological failure can extend across several institutions. An interruption affecting an identity verification system could have implications for access to government services. A failure affecting electronic health records could disrupt the availability of patient information across public hospitals. An interruption in revenue administration could affect tax collection, business transactions and public finances.
The more services depend on shared digital infrastructure, the more important it becomes to understand the dependencies between them. Guyana must therefore move beyond measuring digital transformation primarily in terms of the number of services placed online, the number of communities connected or the speed with which transactions can be completed. It must also measure the resilience of the systems being introduced.
Can essential government services continue when the internet is unavailable? Are government databases backed up independently? Can public agencies restore their systems following a major technical failure? Are there procedures for maintaining essential services when automated systems become unavailable? Can the government identify every external contractor and software provider with access to sensitive public information?
These are practical questions of national governance. The implications for the banking sector are equally important. Guyana is moving towards a more integrated digital payments environment. Th