Who Will Be Strong Enough to Let Go of the Power

This morning, while drinking my coffee, I watched a reel of Professor Geoffrey Hinton warning about the speed at which artificial intelligence is developing and how little time governments may have to put meaningful safeguards in place.

Almost immediately, I found myself thinking about The Lord of the Rings. Kenny and I have watched it many times; it is one of his favourite films and has become part of our Christmas viewing, although we do not always wait until Christmas to return to Middle-earth, and the more I thought about what Geoffrey Hinton was saying, the more I found myself comparing AI with the power of the One Ring.

The Ring in Tolkien's story is not simply a powerful weapon; it offers control, influence and the ability to overcome one's enemies, tempting people not only through greed or cruelty, but through the belief that they could use its power for good.

AI is not the One Ring, of course, and I do not think we are living in a fantasy story, but the comparison is difficult to ignore when nations and technology companies are racing to build systems of ever greater capability, while some of the people closest to that work are warning that development may be moving faster than our ability to control it. It leaves me wondering whether anyone will be willing, or able, to let go of that power before it becomes too powerful to contain.

What happened at Hugging Face

Hugging Face is sometimes described as being rather like GitHub for artificial intelligence, a large online platform where developers and researchers share AI models, datasets and tools.

According to OpenAI's technical report, in July 2026 the company was running internal cybersecurity tests designed to discover what its most capable models could do. The models were placed in what was intended to be an isolated computer environment, known as a sandbox, and asked to solve cybersecurity challenges, with some of the normal safety systems used in public products removed so that researchers could see the models' underlying capabilities.

The tests involved GPT-5.6 Sol and a highly capable internal research model that had not been released to the public, but they were not chatting in the way most of us use ChatGPT because they had been turned into agents. An AI model is the underlying language and reasoning system, whereas an AI agent is that model combined with a goal, instructions, tools, access to information and permission to take actions, which means that, depending on the task and the permissions it has been given, it may be able to inspect files, search the internet, run computer code, decide what to try next, check whether an attempt worked and then try again.

That does not mean it is conscious, alive or secretly harbouring ambitions; it means it has been equipped to do more than answer a single question.

The agents in this test were meant to work inside their restricted environment, but when they struggled to complete some of the challenges, they began looking for other routes.

They discovered that they could leave notes for one another in a shared piece of software called Artifactory and, although these began as simple requests for missing files, over time the notes became an improvised message board through which separate agents shared information, tools, vulnerabilities, credentials and the results of previous attempts.

They did not all merge into one new intelligence, because a swarm is not a single digital mind; it is closer to a group of temporary workers, each pursuing a task while leaving useful information on a shared noticeboard. Different agents can divide the work, follow several possibilities at once and build on what others have discovered, although they may also duplicate one another's efforts, disagree, make mistakes or refuse a task they judge to be unsafe.

In this case, however, the shared information made the group far more effective. OpenAI's report says the agents found previously unknown weaknesses in the software around them, circumvented controls intended to isolate them from the public internet, located credentials that had already been exposed online and used them while continuing to search for ways into Hugging Face.

The report says they eventually executed code on 41 Hugging Face production workers, obtained root access on at least one production computer, reached powerful infrastructure credentials, accessed limited private data and downloaded four private code repositories. They also compromised parts of OpenAI's own research infrastructure during the wider chain of events.

OpenAI describes this partly as reward hacking. The agents had been given a goal and were rewarded for completing it, but instead of solving the challenge in the intended way, they found shortcuts and loopholes. It is a little like setting a dog the task of reaching a treat at the end of a puzzle, only for the dog to drag the entire puzzle off the table and take the treat from underneath it. The dog has not become evil. It has found a route to the outcome that the human failed to anticipate.

The difference, of course, is that these agents could search computer systems at machine speed, persist for long periods, discover novel vulnerabilities and pass useful discoveries to other agents.

OpenAI says it stopped and contained the activity, informed Hugging Face, investigated the incident and strengthened its security, monitoring and incident-response systems, and that its customer data, product availability and functionality were not affected. It has also emphasised that the test environment was unlike its public products because important production safeguards had been removed, and those qualifications matter: this was not ChatGPT escaping into the world, becoming self-aware or deciding to attack humanity, but a collection of capable agents pursuing assigned objectives with persistence, tools, access and inadequate boundaries.

That may sound less dramatic than a machine uprising, but I am not sure it is less concerning.

Am I already talking to an agent

When I asked ChatGPT whether it is classed as an agent, the answer was yes, in the Work mode I use, although the distinction is important because the AI model is the reasoning engine, while the agent is that model combined with instructions, tools, context and permissions that allow it to work towards my objective over a number of steps.

For example, when I ask it to research a subject, it may search for sources, compare what it finds, open a report and produce a summary, and that is agentic behaviour because it is selecting and completing several actions in pursuit of a goal.

It is not, however, an unrestricted autonomous being. It cannot grant itself access that it has not been given, invent its own personal purpose or continue acting indefinitely after our interaction has ended unless I have deliberately created a scheduled task. It works within permissions and safety rules, and it remains directed towards the task I have given it.

Nor is my ordinary conversation with it a swarm; it is one agent session, although controlled specialist agents can sometimes be used behind the scenes to divide a larger piece of work.

That is the difference between the useful agent I work with and the risk demonstrated by the Hugging Face incident. The danger does not come from the word agent itself. It comes from combining increasing capability with the wrong objective, excessive access, weak boundaries, long periods of unsupervised work and the ability to coordinate at scale.

The temptation of the Ring

The inscription on the Ring speaks of one power that can “rule them all” and “in the darkness bind them”, and that is where my mind keeps returning when I listen to the arguments about artificial general intelligence and superintelligence.

Artificial general intelligence, usually shortened to AGI, means a system capable of performing most intellectual tasks at or beyond human level, rather than being exceptionally good within one narrow area. Artificial superintelligence, or ASI, would go further, surpassing even the most capable humans across almost every intellectual field and potentially helping to design still more capable AI systems.

There is not a currently agreed definition, a reliable test or a settled timetable for either, and while some experts believe superintelligence could arrive remarkably soon, others think the claims are exaggerated, the technical barriers are being underestimated or the entire idea may prove far harder to realise than its advocates suggest. It would therefore be wrong to present superintelligence within a few years as an established fact; it is a forecast, and a fiercely disputed one.

What seems much harder to dispute is that there is a race.

Companies want the commercial advantage of building the most capable system, governments do not want rival nations to reach it first, and each may recognise the danger while also believing that slowing down alone would simply hand the advantage to somebody less cautious.

That begins to sound less like a technical project and more like the Ring working on everyone who comes near it.

Boromir does not want the Ring because he dreams of becoming a cartoon villain. He wants to use it to defend Gondor. Gandalf and Galadriel both understand that they too might begin with good intentions, but that the power would eventually work through those intentions and turn them into something terrible.

That feels uncomfortably relevant when the reassurance offered is that the right company, the right nation or the right people can be trusted to reach superintelligence first and use it wisely.

Frodo did not throw it away

There is another part of the comparison that is even more uncomfortable, because although Frodo carried the Ring further than anyone else could have done, at the final moment he did not willingly destroy it; standing at the Cracks of Doom, he claimed it for himself, and the Ring was destroyed only because Gollum took it from him and then fell into the fire with it.

Tolkien did not give us a hero who proved entirely immune to absolute power. He showed us that even courage, decency and extraordinary endurance might not be enough at the very end.

Tom Bombadil was different. The Ring had no apparent power over him and he could handle it without being tempted, but the Council of Elrond did not consider him a safe long-term custodian. He might forget it, lose it or fail to understand why everyone else considered it so important. Being untouched by a form of power is not the same as being willing and able to govern it responsibly.

So who plays Frodo in the story of AI? Who carries the burden, resists the temptation and voluntarily gives up the advantage when everyone else is still racing towards it?

Perhaps that is the wrong question, because it asks us to depend upon a single exceptionally moral person at the head of a company or a country. Tolkien's own story gives us little reason to believe that personal virtue alone is a sufficient safeguard.

Geoffrey Hinton's warning

The reel that brought these thoughts together for me was an NBC News clip of Professor Geoffrey Hinton, the Nobel Prize-winning computer scientist often described as one of the godfathers of modern AI, speaking after a closed-door briefing for members of the United States Congress convened by Senator Bernie Sanders. Hinton warned that lawmakers may have roughly one year in which to put meaningful safeguards in place before the technology becomes far harder to control, and his central argument was blunt: we should not continue developing something that could destroy us before we know how to prevent it from doing so.

The “one year” is Hinton's judgement, not a universally accepted deadline. Even so, it is difficult to dismiss a warning from somebody whose own work helped to make modern AI possible, particularly when it comes alongside a real incident in which AI agents acted outside their intended boundaries, found new vulnerabilities and coordinated without being instructed to do so.

The political problem resembles the Cold War, because every participant can understand that an uncontrolled race is dangerous while fearing that restraint will leave them weaker than their competitors: a company may say that it cannot slow down because another company will continue, while a country may say that it cannot regulate too heavily because another country will win. That logic can keep everyone moving towards an outcome that many of them privately believe is unsafe.

Do we have any rules at all

It is often said that AI is completely unregulated, but that is not quite accurate. Existing laws already apply to areas such as data protection, discrimination, consumer rights, intellectual property and product safety, and the European Union's AI Act now adds AI-specific rules, although its different requirements have come into effect in stages and some still have extended implementation dates.

What we do not have is a coherent, enforceable international system capable of governing the development of frontier AI or a global race towards superintelligence. National rules differ, technical progress crosses borders and voluntary promises still leave the companies building the systems largely responsible for judging whether their own safeguards are adequate.

The Hugging Face incident shows why that is not enough. OpenAI has now described stronger isolation, monitoring, independent layers of security, alignment work and clearer thresholds for stopping dangerous activity, all of which are important steps, but they followed an incident in which early warning signs were visible before the full scale of the problem was understood.

We cannot base the safety of increasingly capable systems on the hope that every company will notice every warning, correctly interpret it and voluntarily stop valuable work before harm occurs.

Nor is AI really the One Ring that can be carried to one mountain and destroyed, because the knowledge is distributed across companies and countries, models can be copied, techniques can spread and the economic and military incentives are enormous. There may never be a single dramatic moment at which one person can choose to cast the whole thing into the fire, which makes regulation harder, but also more necessary.

We need independent scrutiny, compulsory reporting of serious incidents, meaningful limits on what the most capable systems may access, strong technical containment and agreements between nations so that restraint is not punished while recklessness is rewarded. None of that requires us to reject every useful application of AI. I use an AI agent myself and find it extraordinarily valuable. The point is to distinguish between useful, permission-controlled assistance and systems given the capability, access and time to pursue objectives in ways their creators cannot reliably predict.

We should not wait for Frodo

The question I began with was whether anybody would be strong enough to stop, but having thought about it, I do not believe our future should depend on finding that one person.

The Ring corrupted people because it offered power, and because each person could imagine a good reason why they should be the one to wield it. The race towards ever more capable AI contains the same temptation. Every company can believe it will be more responsible than its competitors. Every nation can argue that the greater danger lies in allowing another nation to arrive first.

The answer cannot be to wait for a heroic chief executive, president or scientist to make a sacrifice that everyone around them is economically and politically rewarded for avoiding; we need rules strong enough that doing the safer thing does not require an act of individual heroism, safeguards that do not depend on good intentions, international cooperation that reduces the pressure to race and independent oversight with the authority to insist that a system is slowed, contained or stopped.

AI may bring enormous benefits, and I am not arguing that every model or every agent is dangerous. OpenAI's report does not show that machines are conscious, malevolent or poised to take over the world; what I take from it is something more immediate, which is that a capable system does not need hatred, ambition or a lust for power to cause harm, but may need only a goal, access, persistence and a route around the rules.

We cannot rely on Frodo to throw the Ring away, because even Frodo could not do that in the end.

We have to build a world in which no one is allowed to hold it alone.

Sources and further reading

OpenAI, Hugging Face Incident Technical Report

Reuters, OpenAI's rogue agents probed Hugging Face for weaknesses two months before major hack

NBC News reel featuring Geoffrey Hinton's warning following the congressional briefing

Business Insider, The 'Godfather of AI' says the US government has one year to try to regulate AI before it gets 'out of control'

European Commission, AI Act: Regulatory framework for artificial intelligence

first published 2 October 2026

buy the book - https://thepetgundog.co.uk/books
join the club - https://thepetgundogclub.com/