Artificial Intelligence has done it again! Or in Kenyan parlance, ‘AI imefanya ile kitu ’. A few days ago, Anthropic, the company behind the Claude family of AI models, issued a chilling report of how bad actors had used its AI tools to execute a raft of illicit activities, including surveillance, hacking government data bases and the manufacture of biological weapons.
The report was released just days after a former researcher at the company, Jacob Coxon, resigned in a huff, warning that rapidly advancing AI could kill all humans by the end of this decade. Microsoft co-founder, Bill Gates, had sounded similar alarm less than a week earlier in a highly-publicised essay, where he warned that the world was teetering on the edge of an AI cliff without a clear plan of pulling back.
The Anthropic report which covers the period between December 2025 and August 2026, is disconcerting in the individual cases it documents, which also include high stakes attempts at political engineering in Kenya, but the larger story it tells is even more unsettling.
In the past, expertise has been one of the greatest barriers to power. A person harbouring nefarious intentions such as hacking a government data base, siphoning billions of shillings from a bank, or designing a designing weapons of mass destruction, needed to couple their intention with knowledge, experience, tools, money and, more consequentially, an organisation capable of bringing these capabilities together.
But it is now clear that AI is beginning to upend this convention. In plain language, the Anthropic report is saying that rapidly advancing technologies are just as quickly lowering the price of expertise. This means, that these tools are converting expertise that was once the preserve of highly skilled teams into a service that much less capable actors can easily access and use, to devastating effect.
That is alarming because until now, investigators of cybercrime, for instance, could exude confidence about catching the bad guys simply by looking at the sophistication of an attack. Operations involving extraction of information from complex computer systems generally required considerable technical expertise to execute. Of course, the attribution was not always obvious but the advanced capabilities themselves gave the investigators crucial leads.
That assumption is now becoming less reliable. Anthropic says its year-long investigations revealed multiple instances of AI agents participating directly in operational tasks rather than merely answering an attacker’s questions or making snippets of computer code. The agents conducted reconnaissance, sought and found vulnerabilities, and actively assisted in the exploitation and processing stolen data. In fact, in some cases, humans merely selected targets and supervised the operations as automated systems executed the bulk of the work.
This revelation suggests that AI has now moved from just providing technical advice to the role of an active operator, almost with a ‘mind’ of its own. And that should particularly concern countries such as Kenya, and Africa at large, not least, because of its potential to exacerbate existing vulnerabilities.
First, Africa is digitising rapidly with massive implication for economic growth, but its defensive infrastructure is not keeping up with the pace. Mobile money, digital banking, e-government, online lending and e-commerce are expanding rapidly across the continent and connecting communities that were once disparate isolated entities. Yet these advances have just as rapidly expanded the space of attacks. INTERPOL's 2026 African Cyberthreat Assessment says Africa had more than 1.1 billion mobile subscribers in 2025 and moved more than $1.1 trillion in transactions, while cybercrime losses reported in its assessment more than doubled from $192 million to $484 million since 2024.
The Anthropic finding changes the economics of this vulnerability considerably. A criminal does not necessarily need to understand the architecture of a Ugandan bank, a Kenyan fintech, or a government database to exploit it. AI can now help them understand nearly any system, design or modify tools, and tailor attacks to different environments and systems. Anthropic explicitly warns that complex and highly-encrypted configurations are no longer automatic protections against attacks, because AI makes them easier to understand and infiltrate.
Further, it is now clear that AI will increasingly supercharge precisely the kinds of fraud to which Africa's digital economy is already exposed. Anthropic documented AI-powered dating personas, industrialised scams, cyber operations and other forms of deception. INTERPOL, on its part, had found that an estimated 55 per cent of reported cybercrime across Africa are now AI-enabled.
This is particularly significant because it intersects with Africa’s highly transactional digital personality. Phones increasingly mediate money transfers, loans, government services, employment, commerce and personal relationships. This means an AI-generated WhatsApp message or cloned voice can trigger an immediate financial transaction, and therefore lead to real-world harm.
But perhaps, more consequentially, the Anthropic report has also exposed an asymmetry that may well be the central question for Africa in the ongoing AI governance debate. The report reveals an unsettling asymmetry between the speed at which attackers can acquire AI capability and the response time for state institutions. While hackers and other attackers can now acquire the tools almost instantly, states cannot build investigative capacity, enact laws, train personnel, and create new ways of inter-agency cooperation nearly as quickly. That asymmetry may well be the central question for Africa in the 143-page Anthropic report.
Finally, the report makes concrete the enduring question of information integrity. In Anthropic’s dating-app case, operators created a believable mix of thousands of synthetic identities with real human beings making video calls and interacting on social media. This made the deception increasingly more credible, raising grave implications for situations such as Kenya’s highly charged electoral processes. Imagine a fabricated voice and video of a presidential candidate announcing on the eve of an election that that she has pulled out of the race in favour of an opponent. How about synthetic images of a supposed ethnic attack spreading on a WhatsApp group in a crowded informal settlement?
In sum, Africa seems to have a much larger question to grapple with in the increasing AI divide. The continent no longer faces just a question of accessing the most powerful AI models, important as that is. But it is now increasingly about whether the continent can truly defend itself against these very models.
