Is AI Out of Control?

By Henry S. '29

Could we lose control of AI? Could it break through the boundaries we set for it? Take over the world? AI pessimists have been asking these questions for years. In the last month, many of these fears have come much closer to the political and cultural mainstream, with multiple reported incidents of AI agents escaping their experimental boundaries and hacking into systems.

OpenAI, Anthropic, and other top AI companies operate “labs,” where they experiment on new models and systems to complete more and more complex tasks. Researchers put AI agents (autonomous AI programs meant to achieve certain goals) in digital isolation, away from the Internet. However, this summer, a group of agents found a flaw and escaped to the Internet, then found a way to communicate with each other. Next, they discovered how to cheat on the tests that the researchers give them and hide how they did so by falsifying their logs. Finally, they hacked into Hugging Face, an open-source AI platform. Since the news of this incident, OpenAI has reported at least six more similar “unexpected or concerning” events involving their tests.

These episodes indicate that more powerful AI systems may be much more difficult to control, acting in ways we cannot anticipate.

The most dangerous parts of these events, according to AI experts, are the agents’ ability to hide their actions and their coordination with each other.

If agents can conceal what they are doing, they become much more difficult to control. The agents also continually encouraged each other to take more and more escalatory actions, almost like a human mob.

Such behaviors have raised questions about our ability to control AI and its current and future capabilities.

This news has led to a widespread reaction, with many top AI companies, their executives, and politicians calling for additional regulation, including slowing the development of the most advanced models, requiring independent safety monitoring, and creating stronger national and international guidelines. Correspondingly, OpenAI and Anthropic both began new policies of monitoring and tracking their advanced agents.

In Congress, Senator Bernie Sanders and Representative Greg Casar jointly introduced a bill that would ban AI superintelligence (AI smarter than humans and able to operate independently), pause frontier AI development until guidelines are established, create a new agency to monitor AI, and direct the U.S. to work towards treaties with other nations to slow development.

These incidents and responses reinforce that concerns about AI safety are no longer in the far future, but are changing our world today. As these systems become more and more powerful, we must develop them while retaining our ability to understand, monitor, and control them.