AI agents were used to breach an unnamed North African government technology authority, exposing more than 300,000 national identity records and data from over 500,000 companies, according to analysis published by the Institute for Security Studies.
The ISS assessment, written by consultant and former BBC foreign correspondent Karen Allen, examines Anthropic’s September threat intelligence report through an African security lens — and finds case studies across Kenya, Central African Republic, the DRC, Sudan and Mali.
How the breach worked
Anthropic’s investigation found its agents were used to collect information, conduct “reconnaissance missions” and “build the infrastructure” for the intrusion.
Allen sets that against the scale of African exposure: some 45 governments now run digital systems for service delivery, according to the African Union Development Agency. As more shift to e-government, she argues, exposure grows unless safety controls keep pace.
INTERPOL has warned that 55% of all reported cybercrime across Africa is AI-enabled.
Fully automated attacks have not yet arrived. “We are not seeing fully automated attacks quite yet,” Alpha Group chief information officer Justin Hornsby told ISS — while noting it could happen, and calling for strict rules on which AI models staff use and tighter restrictions on data access.
Surveillance in Mali
Claude models were used in surveillance operations in Mali, with Anthropic finding the agent could breach safeguards and was used to target and profile opposition figures.
Allen notes parallel concerns in South Africa about AI attacks on street cameras and connected devices, with some experts urging the information regulator to review rules struggling to keep pace.
The detection-defeating technique
The most consequential technical finding concerns Central African Republic.
Investigators detected a network of fake radio stations producing AI-generated content — and found operators using Claude to strip away formatting, the telltale signature of synthetic generation.
That defeats detection systems built to recognise those signatures. It is AI used to remove the fingerprints of AI, and it has received little attention in coverage of the report so far.
Investigations into information manipulation in CAR have previously found local agents hired by Russia’s Wagner Group producing pro-Russia content for domestic audiences.
In the DRC, Anthropic’s investigators found 318 articles across several fake news sites, with Claude appearing to have “humanize[d] and refine[d] batches of 50 pre-drafted topics and tweets” for distribution without adequate human moderation.
Kenya 2027
Claude was also used for a domestic campaign, generating batches of 50 tweets made to “look like spontaneous grassroots commentary rather than a coordinated campaign” ahead of Kenya’s elections — including the false narrative that the opposition was “breaking apart.”
Astroturfing has featured in Kenyan elections before. Allen’s point is that AI changes the economics: it accelerates distribution and lowers the barrier to running such a campaign at all.
Where the response falls short
Anthropic says it disrupted most of these operations, while acknowledging that better detection requires intelligence sharing between private sector and government.
Allen identifies two obstacles.
Internationally, the recent Pentagon boycott of Anthropic may signal reluctance among governments and technology companies to cooperate.
Domestically, South Africa’s cybersecurity remains siloed. Threat intelligence passes informally between banks and government branches, but systematic sharing is slow. A private sector-driven Africa Cybersecurity Center of Excellence now exists; the state-led Cybersecurity Co-ordination Centre, promised a decade ago under the National Cybersecurity Policy Framework, does not.
South Africa also has no national AI strategy — the draft was withdrawn in April over fabricated citations, with a redraft due at Cabinet by March 2027. Government agencies invest two-thirds less than corporates in cybersecurity, and the state runs legacy systems that are harder to patch.
Nigeria, Allen notes, has a policy committing it to responsible and ethical AI development.
The starting point
Concerns are not confined to Africa. OpenAI revealed in September that its agents interacted with US government websites in “unexpected ways,” alongside unconfirmed reports they unsuccessfully attempted to access one.
Allen’s recommendation is modest and immediate: basic AI usage policies. Staff and government workers risk feeding sensitive information into public models as they experiment — and in doing so, exposing their organisations to further intrusion.
Based on analysis by Karen Allen for ISS Today.
