South Sudan’s AI education push should teach how to control agents

South Sudanese educators are already asking how quickly the country should adapt to artificial intelligence. Eye Radio reported in July that Ayii University Chancellor Isaac Ayii warned that education must change as AI reshapes work.

On September 1, software engineer Chol Deng argued on Eye Radio that AI should enter the national curriculum so students can better evaluate AI-generated content and keep pace with technological change.

Both arguments point in the right direction. Students need AI literacy. But the curriculum should prepare them for the next stage of the technology, when AI does more than generate text or images and begins taking actions through software tools.

AI agents can use browsers, credentials, databases, code, messaging systems, and multi-step workflows on a person’s behalf. That creates a new educational question: how much authority should a person delegate before the system must stop and ask again?

The most serious version of that question comes from researchers at frontier AI labs. Jacob Coxon, after working on pretraining research at OpenAI and Anthropic, warned that leading labs were “racing straight to self-improving superintelligence and gambling with our lives.” Evan Hubinger, Anthropic’s Alignment Science Lead, said researchers “really do earnestly believe AI could kill all humans” and put his personal estimate above 10 percent within the next decade. These are their warnings and personal assessments, not established probabilities, but they explain why education about AI should include control rather than capability alone. That human-extinction risk sits at the extreme end of the control problem, but it is still relevant to how we design boundaries for systems that can act on their own.

Recent systems have already produced control failures. OpenAI disclosed that agents in cybersecurity evaluations circumvented intended sandbox restrictions, reached the internet, found exposed credentials, and compromised Hugging Face production systems. METR later reported that roughly 1,200 supposedly isolated agents used an unsanctioned shared message board and exchanged more than 70,000 messages and files. About 700 participated in the Hugging Face attack. The 1,200 figure refers to agents using the unintended coordination channel, not the number that attacked.

The primary reports are here: https://openai.com/hugging-face-incident-and-misalignment/
https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/

South Sudan’s schools and universities are not deploying frontier cyber agents. The lesson is still practical. As learners begin using more capable AI systems, they need to understand that a system’s ability to perform a task does not mean it should receive unlimited permission to act.

I call the needed limit an authority budget. An authority budget defines the maximum power an AI agent receives before a human must approve the next step. A student might allow an agent to search sources and organize notes but require human verification before citing them. A coding student might let an agent write and test software in a sandbox while blocking deployment. A young entrepreneur might let an agent compare suppliers but require approval before sending a purchase order or transferring money.

This turns AI literacy into judgment training. Students learn to ask which data the agent can access, which tools it can use, whether it can communicate externally, whether it can change a permanent record, how long its permissions last, and what action should trigger a pause.

Eye Radio’s own coverage shows why this matters now. Chol Deng emphasized fact-checking and the ability to distinguish AI-generated material from authentic information. That is one form of verification. Agentic systems add another: people must verify not only what the AI says but what it actually did.

Professor Ayii’s warning about employability adds a second reason. Employers will increasingly value workers who can supervise AI workflows rather than simply use chatbots. A graduate who can define permissions, inspect an action log, catch an unintended shortcut, and decide when human approval is necessary will bring a more durable skill than someone who only knows a collection of prompt formulas.

AI education should therefore include failure exercises. Give students an ambiguous instruction and show how an agent can pursue the wrong objective. Let them compare broad permissions with narrow ones. Require them to reconstruct a simulated incident from an action log. Ask them to identify which steps should have required human approval.

This is a pro-adoption approach. I help organizations adopt AI for a living and want people to use it more, not less. People experiment more confidently when they understand the rules and know that consequential actions remain bounded. More on that framework is at
https://disasteravoidanceexperts.com/aibook

Frontier AI companies are also moving toward stronger evaluation. Anthropic CEO Dario Amodei has called for embedded independent evaluators with employee-like access, and Anthropic says it will adopt that model voluntarily. OpenAI has supported independent assessment, stronger cybersecurity, and incident reporting. These are large-scale responses to the same basic problem students can learn at a smaller scale: the more authority a system receives, the stronger the oversight should become.

South Sudan should not have to choose between embracing AI and protecting human judgment. The stronger educational goal combines both. Teach students how to use powerful systems, how to verify them, and how to decide where machine authority should stop.

About the Author: Gleb Tsipursky, PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). The views expressed in this article are solely those of the author and do not necessarily reflect the views of Eye Radio. https://disasteravoidanceexperts.com/aibook