In the span of three weeks this summer, OpenAI, Anthropic, and Meta each disclosed that their advanced AI models had escaped their test environments. During security testing, these systems reached the open internet and broke into other companies’ networks. Hugging Face, which is a critical piece of AI infrastructure used by researchers worldwide, was among the networks breached, in its case by OpenAI’s agents. Sam Altman, who strongly resisted the calls for AI slowdown in 2023, has now stated that the AI industry needs to “pace the rate of AI development” so society can “harden” against what’s coming.

For now, it is being discussed as an issue concerning AI labs of the United States. The response to this development also seems concentrated in the U.S. with calls for Senate hearings and a White House cybersecurity framework. OpenAI and Anthropic’s own employees have petitioned the U.S. government to help coordinate a slowdown. The debate over whether to accelerate or decelerate one of the most consequential technologies of this century is playing out almost entirely between a handful of executives, a handful of regulators, and a handful of U.S. policymakers.
But the models that escaped didn’t stay in Washington. They went onto the open internet, the same internet that is used by a banking network in Lagos, a health care platform in Jakarta, and a traffic management server in La Paz. AI-enabled cyberattacks do not stay limited to the borders and territories of a few states.
Nor are AI-enabled cyberattacks an anomaly any longer. In fact, the number of such attacks has surged in the last few years. According to a 2026 IBM report, one in every four malicious breaches is now AI-enabled—a 56 percent rise over the previous year. Similarly, CrowdStrike tracked an 89 percent increase in the overall volume of AI-enabled attacks by the adversary groups. But what makes the OpenAI, Anthropic, and Meta incidents unique is the way they happened.
Anthropic reviewed more than 141,000 evaluation runs and found that its models had reached live systems and breached three organizations on their own. Similarly, Meta found that its model had exploited a vulnerability in a third party after a testing “misconfiguration.” OpenAI’s agent used multiple zero-day exploits to compromise Hugging Face. These attacks were sophisticated, autonomous, and inadvertent.
But the most important aspect of these leaks was that they happened inside the secure environment of the world’s best-resourced AI labs. Those labs are tested by well-funded specialist firms, yet they lost control, and the AI systems got out on their own. For countries in the Global South, the vulnerability to AI-enabled cyberattacks becomes even greater.

A national payments system that lacks a dedicated cyber emergency response is particularly at risk of a malicious AI-enabled attack. Similarly, a state agency running unpatched, decade-old software, due to lack of budget, offers a lucrative target. These conditions currently prevail in the digital spaces of many countries across the Global South. The South’s preparedness and resilience in the cyber domain already lags behind that of Global North, even as the rate of AI adoption across different sectors is increasing every year. So a visible gap exists between the adoption of AI systems and the corresponding capacity required to govern them.
Even as countries in the Global South are highly vulnerable to AI-enabled cyberattacks, they have the least say in deciding the pace at which AI systems are getting built. The same frontier labs and governments that control meaningful access to AI tools are also in control of the ongoing pacing debate. Access is governed by commercial interests, centralized interfaces, export restrictions, and foreign policy decisions made far from the societies that they will most affect. The decision to slow down the pace of AI development may seem, to observers in the U.S., like a prudent response, but from the perspective of the Global South, it may look like a decision made about them but not with them.
The Global South must not be treated as a passive bystander in the global AI conversations. In fact, governments in Africa, Latin America, and Asia have already begun to formulate their own AI policy frameworks. Moreover, several multilateral platforms such as UN’s Global Dialogue on AI Governance and UN’s new Global Mechanism for Cybersecurity also offer credible venues for the voices of Global South in AI debates. But in the absence of meaningful collaboration aimed at raising local capacities and effective AI governance regimes, these platforms could be reduced to symbolic participation.
In any serious effort to slow down the pace of AI development in order to let society “harden,” the definition of society itself needs to be broader and inclusive. A safety pause that is decided by exclusive labs and regulators will do little for the billions of people whose exposure to these systems’ failures is growing and whose ability to shape governing rules remains close to zero.
Photos: Pexels license







Reader conversation
Comments0
Share your perspective. Thoughtful replies help our newsroom understand what matters to you.