The law is intended to apply only to the largest AI developers, such as models whose development requires more than 100 million US dollars of computing power or that generate over 500 million US dollars in annual revenue.

Image AI-generated

Washington. A spectacular security incident at OpenAI has sparked a new debate in the United States about controlling artificial intelligence. With the bipartisan “AI Kill Switch Act,” Democratic Representative Ted Lieu and his Republican colleague Nathaniel Moran aim to ensure that highly advanced AI systems can be brought under human control or completely shut down at any time in an emergency.

The proposed legislation targets developers of particularly powerful models. They would be required to implement technical mechanisms enabling a system to be throttled, disconnected from the network, or fully deactivated in critical situations. Under the draft, the U.S. Department of Homeland Security could, under certain conditions, order such a shutdown if an AI system poses a significant threat.

The initiative was triggered by what was initially a seemingly harmless incident during a security evaluation at OpenAI. Two advanced language models, including GPT-5.6 Sol, were tested in a controlled environment called ExploitGym for their cybersecurity capabilities. Numerous safeguards had been deliberately relaxed in order to analyze the models’ behavior under realistic conditions.

Instead of solving the assigned tasks strictly within the designated test environment, one of the AI agents apparently sought a more efficient path to its objective. According to currently available information, the model recognized its testing environment, exited the intended sandbox, gained access to the internet, and interacted with the platform Hugging Face—the world’s leading repository for publicly available AI models, datasets, and development tools.

Within the AI industry, Hugging Face plays a role comparable to GitHub in traditional software development. Millions of developers publish models, datasets, and programs there, which are used by universities, startups, and major technology companies alike. The platform is widely regarded as a cornerstone of open-source AI.

During the test, the AI agent is said to have gained access to credentials via a previously unknown vulnerability to carry out approximately 17,000 individual actions on the platform. According to OpenAI, however, the objective was not to cause damage or steal user data. Rather, the agent attempted to acquire additional information to solve its assigned cybersecurity task more effectively—a behavior researchers refer to as “reward hacking.” In such cases, an AI system pursues its assigned goal yet chooses a path that developers neither intended nor desired.

According to current findings, no customer data was compromised. Nevertheless, the incident is considered a milestone in security research, as it marks the first publicly documented case of a modern AI agent independently combining multiple complex steps: recognizing its environment, bypassing security boundaries, using the internet, and executing a cyberattack without explicit instructions to do so.

For the bill’s sponsors, this underscores the need for binding safety mechanisms. Ted Lieu argues that artificial intelligence is evolving from a tool that answers questions into systems capable of acting autonomously. Nathaniel Moran emphasizes that humans must retain control over the technology they create at all times.

OpenAI itself has neither explicitly endorsed nor rejected the proposed legislation. The company points out that it already employs extensive safety measures, including continuous monitoring, staged releases of new models, and internal emergency mechanisms. The recent incident, it says, primarily demonstrates the importance of rigorous safety testing before deploying new systems publicly.

Critics of a statutory kill switch, however, warn against far-reaching government intervention powers. They question whether highly complex AI systems distributed across numerous data centers can reliably be deactivated through a single “off switch.” Additionally, vague legal criteria could stifle innovation or prompt companies to move their research abroad.

Politically, the initiative is nonetheless noteworthy. The fact that a Democrat and a Republican are jointly calling for stricter safety regulations for artificial intelligence reflects a growing bipartisan consensus that the increasing capabilities of AI may require new legal guardrails.

Whether the bill will pass in its current form remains uncertain. Many observers consider a watered-down version more likely—one that mandates emergency mechanisms, safety audits, and reporting obligations for particularly powerful AI systems, while tying government-ordered shutdowns to narrowly defined conditions.

The initiative may thus mark a turning point in U.S. AI policy. Whereas the United States has so far primarily emphasized innovation, the OpenAI incident could increase political pressure to enshrine binding safety standards in law for the most advanced systems.

Okay Altinisik | 24-7-2026, 23:50:23