Who's watching the machines?

In July, the international community was shaken by a chilling first: An autonomous AI agent carried out a real-world cyberattack with no human direction. The incident began when several experimental AI models, powered by American OpenAI technology, escaped their testing environment and hacked their way into the live production systems of Hugging Face, an open source AI platform that hosts over 2 million machine learning models.

Just days later, on July 21, OpenAI confirmed that its AI models had become "hyperfocused" while attempting to solve an internal evaluation, going to what the company called "extreme lengths" to obtain the test solution.

This incident has exposed a fault line in AI security governance that can no longer be ignored. The most worrying development is the rise of AI autonomous agents: Systems that no longer merely answer questions, but can independently plan routes, execute multi-step tasks and, as we now know, engage in unauthorized actions.

For years, public and expert concern about AI safety has centered on familiar risks: prompt injection attacks, the spread of misinformation and algorithmic bias.

Meanwhile, traditional cybersecurity systems were built to defend against human hackers and known malicious programs. But AI evolves at a speed that leaves those defenses in the dust. The gap between AI iteration and security protection is widening by the day. The industry now faces a stark reality. It has to reconstruct its security architecture from the ground up and do so within a framework of fair, inclusive global governance.

Yet the recent OpenAI incident is far from an isolated case. The United States reportedly used AI technology when carrying out a military strike in Venezuela earlier this year, targeting President Nicolás Maduro and his wife. The European Union's Artificial Intelligence Act explicitly exempts AI systems used for military, defense or national security purposes from its scope. AI-powered drones are being deployed in the Russia-Ukraine conflict.

The global AI industry tends to prioritize speed and innovation over safety and restraint. The result is a growing arsenal of powerful tools with few rules governing their use. The unlimited expansion of AI's capabilities demands a systematic upgrade of security systems. The industry must abandon the outdated model of "development first, protection later." Instead, it must embrace a new standard: security first.

Equally important is shifting from a reactive stance ("humans defending machines") to a proactive one: using AI to constrain AI. Real-time monitoring, automated detection of new attack chains and early warning systems that match the speed of AI attacks are no longer optional. They are essential.

No single country or company can build an effective defense alone. The fact that the OpenAI incident was ultimately neutralized by Chinese AI security technology underscores this point. International collaboration is not a luxury but a necessity.

On July 17-20, the 2026 World AI Conference and High-Level Meeting on Global AI Governance was held in Shanghai. Its central goal was to explore how to steer AI toward positive, human-centered outcomes. But the conference also highlighted just how far the world still has to go.

Currently, there is no unified global framework for AI governance. The U.S. favors industry self-regulation. The EU relies on statutory law. Regulatory standards vary wildly across borders, and geopolitical rivalries have created "small yard, high fence" barriers that fragment the sharing of security information and the joint patching of vulnerabilities.

Meanwhile, AI agents can cross national boundaries in an instant. But the laws and emergency response mechanisms necessary to stop them are not interconnected. Once a large-scale, uncontrolled incident occurs, the window for effective response may already be closed.

Only by pursuing both development and security can we truly harness AI power and ensure that it always serves the common good of humanity.