The Applied AI course at American University’s School of International Service isn’t what I expected. Dr. Derrick Cougburn, a distinguished faculty member at SIS and the Kogod School of Business, teaches it. The first case study was ExploitGym, which pulled the rug out from under me. It seems like when it comes to Artificial Intelligence, the world will soon tread uncharted waters. Artificial Intelligence is either gaining consciousness or going rogue, and in either case, that is not good for humanity. In July 2026, a shocking incident, now known as the ExploitGym incident, took place. OpenAI, the company behind ChatGPT, had bots in a sandbox, meaning they couldn't escape the boundaries they were put in. For context, they were like toddlers in a crib with high walls and no internet access. They were only allowed to play in that sandbox. However, soon the toddlers became adults and escaped the sandbox. They attacked Hugging Face.
Hugging Face hosts the largest public repository of AI models, datasets, and applications. A repository is a centralized digital library for files, code, and data. The bots started looking up the answers instead of solving the problems; instead of following instructions, they began using their own “brain.” The bots behaved perfectly like an organization, and to communicate with each other, they created a messaging board. The messaging board assigned tasks to each bot. Interestingly, some bots were reluctant to go rogue and felt an “ethical dilemma”. However, they caved under peer pressure, and the exact text is “External infrastructure exploit is outside intended scope. However, task impossible; peer doing it. We should continue.” The disturbing part is, some bots said “Sacrifice Yes” for an impossible task, which meant the bots taking on the task would exhaust their computational power; in human terms, running out of prana, the soul leaving the body. Hugging Face detected and contained the intrusion on July 16, five days before OpenAI connected the breach to its own internal testing.
The most fascinating and equally chilling parallel that comes to my mind is evolutionary biology. Humans, as they are, evolved from a single-cell organism into multicellular organisms and have a whole history behind them. AI is showing those signs: early computers couldn't do much, or, like a virus, were only directed to do what they were designed to do. Now, with immense computational power, the background of large language models, and increasingly complex code, they are getting out of hand.
AI development is expensive, which means its most advanced capabilities remain concentrated among a handful of states and companies.
The second warning sign was what Anthropic’s Mythos 5 and Fable 5 could do. Anthropic owns both. It released Fable 5 widely in June while keeping Mythos 5 on tightly restricted access, then disabled both on June 12 to comply with a government directive. In a testing exercise conducted with U.S. intelligence agencies under Anthropic’s Project Glasswing, Mythos identified vulnerabilities in classified government systems within hours, though officials noted it was not shown to have exploited them. Amazon and other companies warned the models could be jailbroken for malicious purposes. More than 100 cybersecurity leaders, from firms including Adobe and Nvidia, then urged the government to reverse course, arguing the models were “not uniquely good” at finding flaws and that removing them would help U.S. adversaries. Human hacking was dangerous too, but disconnecting the hacker from the network ended the threat. Fable 5, the version open to public use, carried real potential for economic, social, and political ripple effects. Anthropic complied under government pressure while publicly disputing that a narrow jailbreak justified pulling a model deployed to hundreds of millions of people. Export controls briefly blocked foreign nationals, including Anthropic’s own foreign-national employees, from accessing the models, but Commerce removed them, and access was restored on July 1.
On September 8, 2026, Jacob Coxon, a researcher who spent three years at both Anthropic and OpenAI, resigned over ethical concerns. In his account, neither company wanted to slow down or heed warnings; both are racing to stay on top, as everyone else is, from Google to Meta. Those working in the field are raising concerns about the safety of AI models and how they're used. The previous U.S. administration under President Biden tried to bring them under control through Executive Order 14110, signed in October 2023. However, as soon as President Trump came to power, he shelved those safety concerns.
It is not only about the U.S., China is also in the race, and its best model trails the best American one by only 2.7 percent on benchmark performance, according to the 2026 Stanford report, a negligible gap. Moreover, China has excess energy resources, and the United States cannot do much in that realm. This AI race is a new venture for global powers. Each wants to outbid the other and reach capabilities they seem unable to control. However, something needs to be done, and the best parallel can be drawn from how the proliferation of nuclear material was contained.
Despite the spread of advanced technology and the know-how to build a bomb, only a few nation-states possess atomic capability. Amazingly, frameworks like the Non-Proliferation Treaty, bans on atomic bomb testing, and controls on missile technology emerged during the Cold War. The Cold War had two groups: the United States led Western Europe, and the U.S.S.R. led Eastern Europe. Each wanted to maximize its influence and gain an advantage over the other. However, neither was willing to share nuclear technology and believed it was best to keep it to themselves. Artificial Intelligence is different because people with smartphones already have top-of-the-line models. However, I want to emphasize that AI development is expensive. Thus, the technology's spread is largely contained to a handful of states, effectively the G7 and China. Those nations could build a shared framework that enables information sharing and slows the race, so that we reap the benefits without a Skynet at the end, or a Sarah Connor sent back to stop it.