Anthropic

Anthropic’s recent threat intelligence report reveals how malicious actors are using its own artificial intelligence (AI) agents to conduct cyberattacks, surveillance operations and online influence campaigns. Africa’s leaders should be especially vigilant, given the continent’s relatively low digital literacy rates, digital divides and the gap between the pace of AI availability and regulatory lethargy. Citizens are vulnerable to data theft, surveillance at scale and being duped in AI-driven influence campaigns. The Anthropic report focused on malicious uses of its AI agent Claude, which is used by some 30 million people worldwide. It said Claude was harnessed to conduct cyberattacks, surveillance and influence operations. Those responsible included state propaganda institutions, their proxies, politically motivated individuals, and criminal and commercial entities. Case studies included Kenya, Central African Republic, Democratic Republic of the Congo (DRC), Sudan and Mali. AI is being used not simply to create content that deceives users, but as a scaling tool. For example, it enables the rapid generation of multiple social media posts as part of influence campaigns, or the writing of tens of thousands of pieces of malicious code that can outpace swift detection. Regarding traditional cybercrime, Claude was used to scan computer networks for vulnerabilities and identify targets for phishing attacks. Interpol recently warned that 55 per cent of all reported cybercrimes across Africa were AI-enabled. While there is still a degree of human involvement in most intrusions, and ‘we are not seeing fully automated attacks quite yet,’ this could happen, Alpha Group Chief Information Officer Justin Hornsby told ISS Today. He said tighter controls, including strict rules on which AI models an organisation’s staff used and restrictions on data access, were needed. Anthropic’s investigation found that its AI agents were used to collect information, conduct ‘reconnaissance missions’ and ‘build the infrastructure’ to hack into an (unnamed) North African government technology authority. Over 300,000 national identity records and data from over 500,000 companies operating in the country were exposed. The potential vulnerabilities across Africa are vast. Some 45 governments are now running digital systems to improve service delivery, according to the African Union Development Agency. As more countries shift towards e-government, exposure to AI risks will increase unless safety controls are ramped up. Some Claude models have also been used for surveillance operations in Mali, with Anthropic’s investigation revealing the agent could breach safeguards and was used to target and profile opposition figures. Likewise, concerns about AI attacks on South African street cameras and other linked devices are growing. Some experts are calling for the country’s information regulator to review its rules, which are struggling to keep pace with technological change. But Claude’s use has arguably been the most staggering in producing influence campaigns at scale across African countries. Several of these operations have been linked to foreign actors, with Russia a key player. In Central African Republic, investigators detected a web of fake radio stations producing AI-generated content. These malicious actors used Claude to strip away formatting, a telltale sign that material has been synthetically produced. This makes it harder to detect, even by defensive AI systems. Investigations in CAR into information manipulation have found that local agents hired by Russia’s Wagner Group were tasked with producing pro-Russia content for local audiences. Anthropic claims the DRC was part of a network to produce fabricated news. Its investigators found 318 articles across several fake news sites that were created and populated with content produced by the company’s AI agents. It says Claude appeared to have ‘humanize and refine batches of 50 pre-drafted topics and tweets,’ enabling vast quantities of content to be distributed without adequate human content moderation. Claude has also been used for homegrown influence campaigns. Anthropic said Claude was harnessed to generate batches of 50 tweets made to ‘look like spontaneous grassroots commentary rather than a coordinated campaign’ leading up to next year’s Kenyan elections. Among the false narratives was that Kenya’s opposition was ‘breaking apart’ ahead of the polls. Such astroturfing campaigns, which try to mimic organic engagement, have previously featured in Kenya’s elections. AI accelerates distribution, making it easy for people to embark on such campaigns. Like in the other case studies, Claude was used to plan campaigns, choose targets and create content in Kenya. While Anthropic says it was able to disrupt most of these campaigns, it acknowledges the need for better detection and prevention. This requires intelligence sharing between private sector and government partners. Yet the recent Pentagon boycott of Anthropic in the United States (US) may signal a reluctance among governments and tech companies to work together. In countries like South Africa, which leads the continent in indigenous AI innovation, a siloed approach to cybersecurity persists. Threat intelligence may be shared informally between, for example, banks and government branches, but systematic sharing is slow. The country has a new private sector-driven Africa Cybersecurity Center of Excellence, but the state-led Cybersecurity Co-ordination Centre, promised a decade ago as part of the National Cybersecurity Policy Framework, has not materialised. Despite being increasingly vulnerable to cyberattacks, South Africa lacks an AI strategy, and government agencies invest two-thirds less than corporates in cybersecurity. The state also uses out-of-date legacy systems that are harder to fix, according to experts at a recent tech conference. In contrast, Nigeria, which almost matches South Africa in AI development, has a policy that commits it to ensuring responsible ethical AI development and deployment. Governments cannot be complacent. OpenAI’s September revelation that its AI agents interacted with US government websites in ‘unexpected ways’ and unconfirmed reports that the same agents unsuccessfully tried to hack into a US government website have heightened global concerns. Introducing basic AI usage policies in African countries is an important starting point. Employees and government workers not only risk feeding sensitive information into public large language models as they tinker with AI, but could also expose themselves and their organisations to further intrusions. Karen Allen, ISS Consultant This article is republished from ISS Today. Read the original article.

Anthropic

Anthropic report shows AI scaling cyberattacks and influence campaigns in Africa

More Anthropic