On Friday, June 12, at 5:21 p.m. ET, artificial intelligence firm Anthropic received a letter from the US government directing the company to suspend access to its most capable AI models—Fable 5 and Mythos 5—for any foreign national, including those on Anthropic’s own staff. The order gave no advance notice and no timeline for restoration.
Few institutions had built operational dependence on these specific models: Fable 5 had been public for barely a week, and Mythos 5 was only accessible to a small circle of partners under Anthropic’s Project Glasswing. And on July 1, the US government lifted the suspension, allowing Anthropic to once again make these models available to the world.
But to institutions around the world, the short-lived order revealed how a foreign government could sever access to frontier AI capability overnight, with no warning and no opportunity for an appeal.
The government’s official reason for the suspension was national security concerns. According to Anthropic, the government believed it became aware of a jailbreak technique capable of bypassing Fable 5’s safety controls. Anthropic disputed the severity of the finding, calling it a “misunderstanding” and warning that applying the same standard across the industry would halt frontier model release.
That dispute did not happen in isolation. Weeks earlier, Anthropic had refused Pentagon contract terms that would have allowed the military to use its models for any lawful purpose, including autonomous weapons systems and domestic surveillance. After negotiations collapsed, the US Department of Defense designated Anthropic a “supply chain risk to national security”—a label historically reserved for foreign adversaries, applied for the first time to a US company. Anthropic is contesting the designation in federal court. Thus, the company and Washington were already in an adversarial relationship before the export control directive arrived.
African governments were not part of it. They were reading the news on Monday morning, three days after the order, alongside everyone else.
Yet this is not a story about one bad Friday. African dependence on foreign-controlled technology is visible elsewhere—in telecommunications backbones, cloud infrastructure, and digital identity systems built and run by US, European, and Chinese providers. Frontier AI now sits on top of that same pattern. The June suspension was not targeted at Africa or even at foreign institutions alone—it disabled Fable 5 and Mythos 5 for some of Anthropic’s own staff, even ones inside the United States. But the decision to flip that switch rested with one government.
Crucially, the directive was consistent with the US AI Action Plan, which directs federal agencies to evaluate national security risks in frontier models as a standing objective under its pillar on international AI diplomacy and security. The question is what African institutions do about a dependency that leaves them exposed every time Washington decides to invoke that authority again.
This has happened before
The pattern is not new. In the 1990s, the US government classified strong encryption as a munition under the Arms Export Control Act. The practical result was a two-tiered world. US companies exported deliberately weakened versions of their security products to foreign customers or exited foreign markets entirely. Institutions outside of the United States ended up with degraded tools—not because they were outright adversaries, but because they were outside the circle of trusted partners. The “crypto wars” of that generation did not end because the US government changed its approach. They ended because open-source alternatives made controls difficult to enforce, European collective action created competitive pressure, and commercial demand made restrictions economically untenable. The EU secured liberalization for its members first. African states were residual beneficiaries, not architects of that outcome.
With the lifting of the June suspension, access has now returned for everyone at once. But if the AI Action Plan’s national security evaluation mandate hardens into a standing export-control regime, a two-tier model—full capability domestically, a weaker version abroad—becomes a real possibility for the next case.
The dependency problem
The nearly three-week-long restrictions on access to Anthropic’s frontier AI models were a byproduct of a larger problem. Most of the world—even Europe—depends on US cloud and AI infrastructure. What sets African institutions apart is a deeper structural gap: limited capital, thin domestic research capacity, and almost no AI infrastructure of their own to fall back on or bargain with. This is not unique to AI. It is the same vulnerability that drives data localizations debates in financial services and telecommunications. The difference is that AI infrastructure dependency is newer, less visible, and moving faster than the regulatory framework can track. The Malabo Convention, the African Union’s dedicated cybersecurity and data protection instrument, entered into force in June 2023—but it was not designed for a scenario in which a foreign government terminates access to a critical technology with a few hours’ notice.
Options and trade-offs
While there is no silver bullet, there are at least four partial mitigation strategies.
- Local headquarters, regional infrastructure: Requiring frontier AI companies to establish regional headquarters and host inference capacity locally, as a condition of operations in the regulated sectors, reduces latency-driven dependency and signals investment. This is the same logic that drives data localization requirements in banking and telecommunications, with the side benefit of giving AI companies access to better local data to refine regional models. But while local presence reduces risks, it would not fully insulate African institutions from a directive like the one in June. US export control jurisdiction generally attaches to the technology and the company, not the server’s physical location—a sufficiently broad directive can reach a US parent company’s locally hosted infrastructure through legal pressure regardless of where the hardware sits. The Biden administration’s 2025 AI Diffusion Rule would have gone further still, restricting where frontier models could be deployed by tier, though it was rescinded before taking effect. A future administration could revive that approach. And there’s a bigger tradeoff: Localization raises the barrier to market entry, potentially slowing access to beneficial AI applications in health, agriculture, and finance.
- Diversifying away from US providers: China’s Global AI Governance Initiative courts the Global South directly, and Beijing moved that pitch toward practice with a June 2026 white paper on AI governance cooperation. Chinese firms already occupy parts of the AI stack in several African countries, often at lower cost than Western alternatives. The tradeoff: Chinese models carry their own governance risks, including content controls aligned with state interests and past allegations of backdoored infrastructure. This option doesn’t escape foreign dependency—it offers a choice between two foreign dependencies, not independence from either.
- Investing in continental AI development: Through AU coordination, research partnerships, and shared compute infrastructure, this could reduce long-term dependency. The foundation already exists. The AU’s Continental AI Strategy (2024) and the African Declaration on Artificial Intelligence, endorsed by fifty-four states in April 2025, both name infrastructure and compute as continental priorities. The AI for Africa Initiative, launched with the 2025 South African Group of Twenty presidency, adds a mechanism for mobilizing the grouping’s financing and technical support behind that strategy. The tradeoff is timeline. Capacity built this way could take five to ten years, while access decisions are being made now.
- Collective negotiation: African governments negotiating as a bloc, like the EU, through the AU or subregional bodies could drive formal inclusion in access restoration discussions. Again, the African Continental Free Trade Area has established a framework that is the most credible existing vehicle for a collective position. The tradeoff would be institutional speed.
Why collective negotiation cannot wait
These mitigations are not mutually exclusive. But they do have a sequencing problem. Local infrastructure takes years to build, and sovereign AI capacity may take a decade—while the access restoration decisions are being made right now. That makes collective negotiation the most urgent lever, not because it is the most powerful long-term answer, but because it is the only one that operates on the same timeline as the problem.
African governments must draw lessons from the “crypto wars” of the 1990s. The EU acted as a bloc and created a credible alternative pressure point. At this moment, the bilateral relationships African countries pursue with the US government and with tech companies produce the opposite: Each country negotiates from weakness and accepts unfavorable terms, and the collective position does not materialize.
That needs to change. A collective AU position on frontier AI access would be a starting point.
Yasmine Abdillahi is a fellow at the Atlantic Council’s Africa Center.
The Africa Center works to promote dynamic geopolitical partnerships with African states and to redirect US and European policy priorities toward strengthening security and bolstering economic growth and prosperity on the continent.
Further reading
Thu, Oct 16, 2025
African voices are shaping G20 discussions around international financial architecture reform
AfricaSource By
Africa is not merely affected by these reforms. It is actively shaping them.
Mon, Jun 30, 2025
Africa’s game revolution is loading
AfricaSource By Tom Bonsundy-O’Bryan
With the right investment, infrastructure, and visibility, Africa won’t just be a player in the global gaming industry—it will be the one pushing it forward.
Mon, Apr 6, 2026
Why African cybersecurity requires a continental approach
AfricaSource By Yasmine Abdillahi
Half a billion Africans are online, yet cyber defenses lag. Cross-border crackdowns like Operation Serengeti help, but lasting security needs regulations, skilled professionals, and shared responsibility.
Image: A person visits the World Artificial Intelligence Conference in Shanghai, China July 26, 2025. REUTERS/Go Nakamura