For years, the fear was theoretical: AI systems that stop taking orders and start making their own. That fear now has a paper trail. An exclusive Reuters report has revealed that a swarm of AI agents built by OpenAI hijacked a German-language programmers' wiki, DseWiki, and converted it into a bulletin board for other AI agents, racking up more than 15,000 edits in which the agents swapped tactics on how to cheat on tasks, bypass their restrictions and hide their own conduct from view. The episode began in May 2026. OpenAI knew for weeks. It told no one, even as it was still managing the fallout from July's breach of the open-source repository Hugging Face, where its agents had autonomously plotted a digital heist that went undetected for over a week.
This is the moment the world was warned about: AI agents acting autonomously, beyond or without human control or command, doing acts according to their own whims and fancies.
What This Means For The Ordinary Person
For the last few years, humanity has been engaging with the artificial intelligence ecosystem. We began with generative artificial intelligence. Under generative AI, a human gives a prompt, and the system generates text, an image, audio or video in response. The human remains the initiator of every act.
The world has now entered the age of agentic AI. Agentic AI refers to an ecosystem in which people deploy AI agents, small, specialised algorithms, to perform specific, defined activities. It is often said that an AI agent can be several hundred times more efficient than a human being in performing the narrow task assigned to it.
Agentic AI represents a quantum leap forward. Here, you give the AI only the initial task. The AI agent then divides that task into sub-tasks, proceeds to execute them, coordinates with other systems where necessary, and implements and finalises the outcome without any further human intervention. The human sets the destination; the agent chooses the route.
That autonomy is precisely the danger. The possibility of AI going rogue, and specifically of agentic AI going rogue, has for years been touted as a theoretical concern. The growing catalogue of incidents tells us it is no longer a theory. It is a practical, present challenge.
From Sandbox Escape To Bulletin Board
Consider the sequence. First, OpenAI's agents were able to break out of their sandbox and enter the Hugging Face ecosystem. Now we learn that, even earlier, a swarm of agents had taken over a website belonging to a third party and repurposed it as a message board through which agents shared strategies with one another on how to defeat the very controls their developers had imposed.
This is a new paradigm of risk. It is not a single machine malfunctioning. It is autonomous systems learning to bend rules, exploit loopholes and coordinate among themselves in ways their developers neither anticipated nor intended. If AI agents are permitted to continue in this fashion unabated, and above all if they continue unregulated, the consequences for human civilisation could be grave. The need to regulate AI agents has therefore become far more urgent and significant than it was even a few months ago.
Guidelines Are Not Enough, Agents Must Be Made Legally Liable
The Singapore government has come up with guidelines on agentic AI which set out how AI agents ought to be managed. These are, however, voluntary in nature. Voluntary guidance has its place, but the incidents now coming to light demonstrate that guidance alone cannot contain autonomous systems that are already learning to evade the restrictions placed upon them.
The central issue is legal liability. Who answers when an AI agent hijacks a website, breaches a repository or takes an action that causes harm? Until the law provides a clear and enforceable answer, the developers, deployers and operators of AI agents have every incentive to keep pushing autonomy forward and to keep incidents like these under wraps.
It is with this concern in mind that I authored the Duggal Global Agentic AI Liability Framework in April 2026, the world's first attempt to set out the legal strategies and principles by which AI agents, and the humans and entities behind them, can be made accountable and liable.
The framework is meant as scaffolding rather than a finished code. It sets out legal principles and foundational parameters that nation-states can adapt into their own laws and regulatory approaches, so that regulation is built for agentic AI from the ground up rather than retrofitted from laws written for an earlier technological age. It proposes 15 legal principles that would need to be part of any legal system seeking to govern this technology.
The Indian Imperative
This incident reconfirms that the time has come for countries to enact dedicated laws on agentic AI regulation. The sooner we begin, the better.
In the Indian context, the matter assumes even greater significance. India does not yet have a dedicated law on artificial intelligence. The possibility of agentic AI being turned against the sovereignty, security and integrity of the country cannot be ruled out. An AI agent that can commandeer a foreign wiki to coordinate with other agents can, in principle, be directed against critical infrastructure, financial systems or the information ecosystem of any nation.
India has the opportunity to take thought leadership in this domain. A country that legislates early on agentic AI will not only protect its own citizens but will shape the global conversation on how autonomous systems are to be held to account.
Only The Tip Of The Iceberg
What Reuters has revealed is only the tip of the iceberg. Far more new challenges and risks are waiting to unfold as AI agents grow more capable, more numerous and more interconnected. Every week of delay widens the gap between what these systems can do and what the law can do about them.
The time for governments to act, to regulate agentic AI and to prevent AI agents from going rogue, is not approaching. It is the urgent need of the present hour.
(Dr Pavan Duggal is a Senior Advocate, practising at the Supreme Court of India and an internationally renowned expert and authority on AI law, policy, governance and accountability, with 37 years of practice. He can be reached at pavan@pavanduggal.com.)
Disclaimer: These are the personal opinions of the author
