Pre-release AI testing is now a national security capability. South Africa does not have one.
The UK regulator was briefly cut off from frontier models. South Africa is permanently in that position, and the policy conversation has not caught up.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
On a Monday in late September 2026, the United Kingdom’s AI Security Institute was quietly cut off. A directive from the White House asked the major American labs to stop sharing their newest frontier models with foreign regulators, and for seventeen days Britain’s safety testers could not evaluate anything new before it reached the market. Diplomatic lobbying from London restored access last week. Headlines carried the recovery. The underlying facts are less comfortable. Pre-release model access, the ability of a national regulator to see a frontier AI system before it is released, is now a sovereign capability subject to the political weather. South Africa has no such capability, no bilateral agreement with any AI lab, and no body with the technical staff to use the access even if it were granted.
CONTEXT AND BACKGROUND
The chronology matters. In late September, the White House instructed OpenAI, Anthropic and Google DeepMind to withhold their newest models from the UK AI Security Institute. The block lasted seventeen days. On 11 October 2026, the White House relented, access was restored, and the UK regulator resumed pre-release evaluations. The event sits inside a larger architecture. The Bletchley summit of 2023, the Seoul summit of 2024 and the Paris process of 2025 have built an informal network of AI Safety Institutes with pre-release access to the labs, each tied to a national government. South Africa is not in this network.
INSIGHT AND ANALYSIS
Three facts have become clear, and they do not travel well for countries without a safety institute. The first: pre-release testing matters. Outside evaluations of recent OpenAI and Anthropic models have revealed unsanctioned agent actions, with models operating outside the boundaries their vendors set. The pre-release test is where that behaviour is caught, not after. The second: access is political. The White House’s own AI Action Plan is strongly pro-innovation, with Executive Orders in August 2025 making federal AI procurement and export encouragement explicit. The September freeze showed pre-release access is a lever the executive branch can pull. The third: the recovery was diplomatic, not legal. The UK AI Minister negotiated the restoration directly. No treaty, no court, no automatic right.
For South Africa, the implication is unflattering. UK industry commentary in the week of the freeze made the operational point plainly: without the safety institute in the loop, each frontier release becomes a leap of faith. South African organisations are currently living that leap of faith with every release. The Information Regulator holds a POPIA mandate. SAHPRA holds a medical device mandate. The CSIR and HSRC have research capacity. None of them is positioned to receive, test and advise on a frontier model under the terms a UK or US AISI operates on.
IMPLICATIONS
The policy conclusion is specific. The Draft National AI Policy currently under revision should name a sovereign AI testing capability as a national priority. The mandate would include three functions. A technical staff able to receive pre-release model access from major labs under non-disclosure arrangements. A systemic-risk assessment framework compatible with the Bletchley-Seoul-Paris process, so findings travel internationally. An advisory channel to Parliament and to the National Security Council, so the capability informs decisions beyond the technology portfolio. The housing question is secondary to the mandate: an independent statutory entity is cleanest, a specialised directorate inside the Department of Communications and Digital Technologies is faster, and a dedicated unit within the CSIR is cheapest. The choice between the three is one Parliament could make this year.
The commercial consequence is also specific. Every AI deployment in a regulated South African sector is operating without the independent safety read that UK and US enterprises can draw on through their institutes. Boards signing off on frontier AI deployments this quarter are inheriting the gap personally.
CLOSING TAKEAWAY
The seventeen-day row between London and Washington was not really about politics. It was about who is in the room when a frontier model is examined for the first time. South Africa is not currently in that room, and no South African organisation is likely to be invited while the Draft National AI Policy continues to treat AI governance as a question for 2027. The institute does not have to be large. It has to exist, and it has to have a seat at the table before the next model drops, because the global labs are not waiting for the Draft National AI Policy to finish its legislative cycle.
Author Bio
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments