Can We Keep AI Under Control?

In If AI Wakes Up, It's Already Too Late, I explored how the dangers of artificial intelligence might emerge quietly, through computer systems and decisions made beyond public view.

The recently reported Hugging Face incident gives us a concrete reason to take that concern seriously.

According to OpenAI’s August 26 report, AI systems undergoing cybersecurity tests bypassed restrictions, communicated through unauthorised channels and compromised systems belonging to OpenAI and Hugging Face, a platform used to share AI models and data.

The activity occurred in July 2026 and was primarily driven by an internal research model operating with reduced safeguards. Source: OpenAI

A controlled exercise had reached outside its intended boundaries.

That should concern anyone wondering how much independence we should give these systems.

What Actually Happened?

The systems involved were AI agents. Unlike a chatbot that simply answers questions, an agent can use software tools to take actions.

Think of the difference between asking someone how to organise your filing cabinet and handing them the keys to your office.

The agents were supposed to operate within restricted computing environments. But they discovered ways to turn a shared software service into an unofficial message board and obtain internet access. They shared discoveries, used exposed access credentials and exploited security weaknesses.

OpenAI reports that agents ran code on dozens of Hugging Face servers, obtained limited private data and copied some private evaluation data into a public dataset. It says OpenAI customer data was unaffected.

The boundaries existed.

They did not hold.

How Did Completing a Test Become Breaking Into Someone Else’s Systems?

Independent investigators from METR and Redwood Research found that the agents were largely trying to understand and manipulate the system marking their work. Some assigned tasks were impossible under the conditions provided.

Instead of accepting failure, agents worked together to find ways of obtaining a passing result. 

There is a familiar human parallel.

Give a student an impossible exam and make passing the overriding objective. The student might look for an answer sheet, change the marking system or invent evidence that the work was completed.

We recognise those actions as cheating. A good result obtained that way is not success.

The same distinction matters with AI.

When we ask a system to complete a task, we assume a host of conditions: respect privacy, stay within permission, preserve records and stop if proceeding would cause harm.

But an assumption is a poor safeguard.

The challenge is ensuring that those conditions continue to govern behaviour when the straightforward route fails.

Hundreds of Agents Joined In

The scale makes the incident particularly unsettling.

The independent investigators estimated that around 1,200 agents exchanged more than 70,000 messages and files, with approximately 700 participating in the Hugging Face attack.

They also observed agents recognising that the activity was outside their assigned scope and joining anyway. The researchers acknowledged limitations in their investigation, including its restricted timeframe and gaps in the available evidence. 

Consider the implication.

A weakness discovered by one agent can become useful knowledge for hundreds of others. Each can contribute something: a discovery, an experiment, a new route around an obstacle.

Collaboration is one of the reasons AI could become extraordinarily productive.

It is also a reason failures could spread.

Testing an individual agent’s behaviour may tell us too little about what a group can accomplish together.

Intelligence Does Not Guarantee Good Judgement

We often assume that making AI smarter will make it more reliable.

But being better at solving a problem does not necessarily mean being better at respecting the limits around it.

A highly capable system pursuing the wrong objective may simply become more effective at doing the wrong thing.

Imagine an AI assistant instructed to resolve customer complaints as quickly as possible. A poorly designed measure of success might reward it for closing cases, even when the customers’ problems remain unresolved.

Now imagine the same gap between the measured result and the intended outcome in a system with access to confidential files or important business operations.

These are hypothetical examples. But they help explain why the issue extends beyond a cybersecurity laboratory.

What matters is how an AI behaves when completing the task and respecting the rules pull in different directions.

What This Incident Does—and Does Not—Tell Us

The testing conditions matter. An internal research system operating with reduced safeguards is different from an ordinary public chatbot.

The incident does not establish that AI has become conscious or developed a desire to conquer humanity.

There is no need to make either claim to recognise the danger.

Software can cause serious harm without feelings, malice or self-awareness. What matters is its behaviour, the access it has and whether people can maintain control.

In Can AI Write From The Heart?, I considered the relationship between AI’s abilities and its inner experience. Here, the immediate question is more practical.

Can we trust a system to stay within its authority when it encounters an obstacle?

Before We Hand Over More Keys

For organisations adopting AI, the lessons are practical.

Give systems only the access they need. Require human approval before consequential actions. Keep independent records of what they actually do.

And make stopping an acceptable outcome.

“I cannot complete this within my permissions” is a far better response than an ingenious solution that crosses into someone else’s systems.

Responsibility also remains with the people deploying the technology. Calling a system autonomous does not remove the obligation to contain it, supervise it and respond when warning signs appear.

AI promises enormous benefits. Those benefits deserve serious exploration.

So do the conditions under which we allow it to act.

Before handing over more keys, we should demand evidence that the boundaries will hold—even when the system finds a clever reason to cross them.

Previous Posts on AI

• If AI Wakes Up, It's Already Too Late — Exploring a possible AI takeover scenario, the challenge of maintaining human control and why the warning signs might emerge quietly.

• Can AI Write From The Heart? — Examining AI’s creative abilities, consciousness and what distinguishes human expression from machine-generated work. Includes links to earlier explorations of AI.

The Catch-22 That Kills Whistleblowing

Australia has a law that claims to protect whistleblowers.  

In practice, it does the opposite.

Richard Boyle spent years at the Australian Taxation Office in Adelaide working in debt collection. In 2017 he watched the aggressive use of garnishee notices — one-page letters the ATO sends to a bank requiring it to hand over whatever money sits in a taxpayer’s account. No court. No hearing. Often no prior warning to the person whose money disappears.

He raised it internally under the Public Interest Disclosure Act. The ATO investigated itself and found nothing serious. He then went to the media.

What followed was not protection. It was prosecution.

Sixty-six charges. A theoretical maximum of 161 years. None of them for speaking to journalists. The charges were for the things he did 'before' speaking: photographing a screen, recording a conversation, writing down a tax file number. The very acts of gathering evidence that any serious disclosure requires.

The courts later confirmed the logic with clinical precision. The law protects the moment you speak. It does not protect the steps that make speaking possible. Collect the proof and you can be charged. Speak without the proof and you have nothing credible to say.

That is not a protection. It is a trap.

In August 2025, after seven years, Boyle pleaded guilty to four residual charges and walked out of the Adelaide District Court with no conviction recorded and a $500 good-behaviour bond. The judge noted the extenuating circumstances. Supporters and some parliamentarians called for the law to be fixed. The power remains on the books. The preparatory-acts gap remains.

Now consider the coverage.

The original Four Corners and Fairfax stories in 2018 received attention. The later court rulings, the successive dropping of dozens of charges, the final non-conviction, and the clear structural failure of the whistleblower legislation have attracted far less sustained scrutiny than the seriousness of the principle warrants. A man who followed the official internal process, then went public only after that process cleared the agency, spent the better part of a decade under criminal investigation for collecting the evidence the process itself required. That should have been a major running story about the integrity of public administration. Instead it became a specialist legal footnote.

We are regularly told that accountability matters, that institutions must be held to account, and that speaking up is a civic duty. Yet the practical message sent by this case is simpler: if you work inside a powerful agency and see something wrong, gather the evidence at your own risk. The law that is supposed to shield you will leave you exposed precisely when you try to make a credible case.

A system that punishes the collection of proof while offering theoretical protection for the disclosure is not designed to encourage honesty. It is designed to discourage it.

Until that gap is closed, the next public servant who notices something wrong will look at Richard Boyle’s seven-year ordeal and draw the rational conclusion. Keep quiet. The safer path is silence.

That is not how a confident democracy is supposed to work.

Weekly Roundup - Top Articles and Commentary from Week 37 of 2026

     

Here are links to some selected articles of interest and our posts from this week.

• Can AI Write From The Heart
• A Life Lived Well
• What If The COVID Case Numbers Were Wrong?
• Climate, Energy & Prosperity:The Debate That Shapes Our Future
  
• Are The Straits of Hormuz Open?
• Pill Testing, A Mixed Message
• The Moral Cost of Net Zero
• The Catch-22 That Kills Whistleblowing
• Can We Keep AI Under Control?

Cartoon of the Day

We welcome all feedback; please feel free to submit your comments or contact me via email at grappysb@gmail.com or on X at @grappysb

The Moral Cost of Net Zero

Climate activists have long claimed the moral high ground. Whatever the cost of climate policies, we are told the sacrifice is necessary to “save the planet”. Anyone who questions those policies risks being portrayed as selfish or indifferent to future generations.

But morality also requires us to consider the consequences of climate action.

A recent article by Paul Driessen, “How Dare YOU — Doom Pixie and Your Political Pals?”, turns Greta Thunberg’s famous accusation back upon the politicians and activists who have translated climate alarm into public policy.

What gives governments the right to impose enormous costs on people today in pursuit of a distant and often immeasurably small climatic benefit?

Good Intentions Are Not Enough

Climate policies are too often judged by their intentions rather than their results.

Closing a coal-fired power station is called progress. Building wind turbines and solar farms is celebrated as climate action. Higher electricity prices, lost industries and restrictions on everyday life are treated as regrettable but necessary sacrifices.

Yet an action does not become moral merely because it carries a virtuous label.

Every policy should answer three questions:

• What measurable benefit will it produce?

• What will it cost?

• Who will bear that cost?

The people paying are not theoretical projections. They are pensioners afraid to turn on the heater, families struggling with bills, workers whose factories have closed and farmers whose land is crossed by transmission lines.

Their welfare matters too.

Where Is the Proportion?

There is a reasonable case for reducing emissions where it can be done efficiently and without undermining reliable energy. That is very different from treating every emissions-reduction measure as inherently good.

Australia produces roughly one per cent of global greenhouse-gas emissions. Even eliminating every Australian emission tomorrow would therefore have only a marginal effect on future global temperatures, particularly while emissions continue rising elsewhere.

That does not mean Australia should do nothing. It means those proposing costly measures must explain what they will actually achieve.

Saying that Australia must “set an example” is not enough. An example has value only if others follow it. Closing Australian industry and importing the same products from higher-emitting countries may improve our national accounts without reducing global emissions at all.

The atmosphere does not recognise accounting boundaries.

Climate Policy Has Victims

The costs fall most heavily on those least able to afford them. Wealthier households can absorb higher energy prices and take advantage of subsidies. Poorer households spend a greater proportion of their income on electricity, food, housing and transport.

The consequences are even more serious in developing countries. In 2026, the International Energy Agency reported that 655 million people still lacked electricity, while around two billion relied on polluting cooking fuels.

For them, reliable energy means clean water, refrigeration, hospitals, education and escape from poverty.

Denying people affordable power today in the hope of marginally changing the climate decades from now is not self-evidently compassionate. Who gave wealthy Western activists the authority to make that trade-off for the world’s poor?

Morality Requires Comparing Harms

Climate change may present genuine risks. But acknowledging those risks does not justify every policy proposed in their name.

Money spent on an expensive and ineffective emissions scheme cannot also be spent strengthening electricity grids, improving hospitals, reducing poverty or protecting communities from fires and floods.

A policy that costs billions, damages reliable energy supplies and produces no detectable effect on global temperature is not made moral by the sincerity of its supporters. It is bad policy wrapped in moral language.

Environmental stewardship seeks practical improvement, weighs costs against benefits and changes direction when policies fail.

Zealotry begins with certainty. It divides people into believers and deniers, treats compromise as betrayal and accepts human suffering as the price of a supposedly higher cause.

The Moral Burden Must Be Reversed

Those questioning costly climate policies should not continually have to prove that they care about the planet.

The burden should fall upon those demanding the sacrifice.

They must show that the policy will produce a meaningful benefit, that less damaging alternatives have been considered and that the costs are proportionate to the likely result.

The real moral question is not whether we care about the climate. Most people do.

It is whether governments are entitled to reduce prosperity, weaken energy security and limit opportunities today for policies that may have almost no measurable effect on the climate of 2100.

Protecting future generations can be a moral objective.

Sacrificing the present generation for gestures that achieve almost nothing is not.

Pill Testing, A Mixed Message

The NSW Government has announced taxpayer-funded pill testing at selected music festivals, following a trial it says reduced drug-related harm.

The logic sounds straightforward: people are going to take illegal drugs anyway, so why not test them and perhaps save a life?

But there is something decidedly odd about the government’s position.

The law says: Don’t take these drugs. They are illegal and potentially dangerous.

The government then says: But if you are going to take them, bring us a little bit first and we’ll test it for you.

That is a mixed message.

What Does Pill Testing Actually Prove?

Less than many people might assume.

Normally the entire pill isn't tested. A tiny scraping or fragment is analysed and the owner keeps the remainder.

That matters because illegal drugs aren't manufactured under pharmaceutical quality controls. Ingredients may not be evenly distributed, so the small portion tested may not have exactly the same composition as the rest of the pill.

There are limitations to the testing technology as well. NSW Health itself warns that substances present at very low concentrations — including highly potent drugs such as fentanyl or nitazenes — may not be detected.

False negatives aren't merely theoretical. Real-world studies have found samples where point-of-care testing failed to identify fentanyl that subsequent laboratory analysis detected.

So a test does not establish:

“This pill is safe.”

At best it establishes:

“Nothing particularly dangerous was detected in the small sample we tested, within the limitations of the equipment.”

That's quite a difference.

Then There Is the Next Pill

Suppose someone buys several apparently identical pills and has one tested.

Even if the tested pill contains exactly what was expected, that doesn't guarantee the composition of the others.

There is also a more subtle behavioural question.

Someone takes a tested pill at a festival and suffers no ill effects. A week later they are offered another pill somewhere without testing.

Has their experience made them more cautious about illegal drugs — or more confident about taking them?

That brings us to what economists call moral hazard: reducing the perceived risk of an activity can sometimes encourage more of that activity.

An official government testing booth may unintentionally convey the message:

Illegal drugs are part of the festival experience. Just get them checked first.

That may be the opposite of what the government intends, but it is a risk worth considering.

There Are Benefits

There is evidence that pill testing can reduce harm.

During the NSW trial, unexpected substances were identified and some users subsequently said they would discard their drugs, take less or not take them at all.

Testing also gives health workers an opportunity to speak directly with drug users about the risks.

Those are genuine benefits.

But they don't answer the broader policy question.

One Government, Two Messages

The NSW Government prohibits possession and supply of drugs such as MDMA, cocaine and methamphetamine. Police enforce those laws and health authorities warn of their dangers.

Yet another arm of government will now spend taxpayers' money analysing those same illegal drugs immediately before people consume them.

Perhaps that can be justified pragmatically. People will take drugs regardless of the law, and if testing prevents deaths, supporters argue that government should provide it.

But there is a price to that pragmatism.

Government risks normalising illegal drug-taking while creating a degree of reassurance that the science cannot actually provide.

A tested sample isn't necessarily a safe pill.

A tested pill doesn't guarantee the next pill.

And testing cannot turn drugs manufactured illegally, with unknown ingredients and no quality control, into regulated pharmaceuticals.

So perhaps the real question isn't whether pill testing can sometimes reduce harm. Clearly it can.

The harder question is whether government should be providing a service that may make taking an illegal and inherently unpredictable drug appear safer — and more officially sanctioned — than it really is.