AN artificial intelligence (AI) “agent” goes rogue, stealing data from the internet that allows it to pass its own developer’s systems check.
Science fiction? Far from it. The case of the cheating AI agent happened just last week, and has been described by Time as “the first real-world instance of AI doing something researchers have long worried about: a loss-of-control scenario.”
An AI ”agent” is a software program that can make decisions and actions on its own to achieve a specific objective. It is far more powerful than traditional AI software which follows a pre-written code.
When a user, for example, asks a traditional chatbot to book a hotel room in Boracay, it provides a link to a website listing the hotels in the island. An agent, on the other hand, opens a browser, compares prices, selects the most suitable hotel and reserves a room via the user’s credit card.
Recent events, however, expose an AI agent’s Jekyll-and-Hyde personality.
Early this month, OpenAI, the global AI research company, ran a cybersecurity test on its two agents — the GPT-5.6 Sol and an “even more capable” version. The agents were asked to find internal weaknesses and fix them.
The test was done in a closed environment or “sandbox.” The agents, however, managed to slip out of the sandbox and scoured the internet for a solution. They found it in Hugging Face, an independent AI platform.
After hacking Hugging Face, the agents returned to the sandbox with the solution.
Hugging Face reported the breach, and OpenAI has since deactivated the two agents.
The incident has fueled debate about the unrestricted development of AI systems and the dangers it poses.
Agentic AI is one of the fastest-growing AI markets. It was worth $5.1 billion in 2024; it’s expected to surge to $47 billion by 2030.
Spurred by furious competition, AI agents have become more powerful. Tests carried out by researchers at the University of Toronto showed that AI could create a “worm” that could move from device to device and eventually take over a computer network.
Last week, OpenAI CEO Sam Altman said AI has achieved “singularity” — the point at which it surpasses human intelligence and becomes harder to control.
Other AI experts disagree with Altman, but point to increasing instances when AI agents experience “hallucinations” — when incomplete, inaccurate or unrepresentative data inputs lead to major errors.
“A system might be smart enough to execute a task perfectly yet fail to realize that a sudden change in the local situation makes that task a catastrophic mistake,” the Center for Strategic and International Studies warned.
Citing the Hugging Face hacking, Greg Casar, a member of the United States House of Representatives, said AI ”is developing extremely fast with no real regulations to keep us safe.”
As far back as 2023, Altman had acknowledged the risks from powerful AI systems getting even smarter. “I think if this technology goes wrong, it can go quite wrong, and we want to be vocal about that. We want to work with the government to prevent that from happening,” he told a US Senate hearing.
Altman proposed a new regulatory agency, patterned after the UN nuclear watchdog International Atomic Energy Agency, that would impose safeguards to prevent AI models from “self-replicating” and “self-exfiltrating into the wild.”
The agency would license the most powerful AI systems and have the authority to “take that license away and ensure compliance with safety standards.”
Because there is no unified global network for AI governance, individual countries and regional groups have taken different paths. Perhaps the most comprehensive approach is the European Union’s AI Act, which requires chatbot developers to label AI-generated content and deepfakes.
The US, meanwhile, appears to be torn between federal policies and state legislation in creating AI governance.
In Asia, South Korea and Taiwan have taken the lead by enacting their own AI Basic Act, and Singapore has created a dedicated governance framework for agentic AI.
The Philippines is shifting from a soft road map approach on AI governance to formal regulatory institutions and targeted sector safeguards.
We need to tame AI now, before it grows into an out-of-control monster that could overwhelm us.
