Is AI Really Getting Out of Control? What Sam Altman’s Warning Actually Means

By Joe Foley | AI for Ordinary People

If you saw the headlines this week, you might think the people building artificial intelligence suddenly realized they created Frankenstein.

OpenAI CEO Sam Altman warned that the rapid development of AI could go “very badly.”

Anthropic CEO Dario Amodei wants AI companies to slow down.

Researchers are talking about AI systems getting around safeguards.

Politicians are debating whether government needs to step in.

And, inevitably, some people are asking the biggest question of all:

Is artificial intelligence actually becoming dangerous enough to threaten humanity?

It is a fair question.

It is also a question that deserves a better answer than either “AI is going to kill everybody” or “there is absolutely nothing to worry about.”

The evidence supports neither extreme.

So let's look at what is actually happening.

What Sam Altman Actually Warned About

According to reporting from The Wall Street Journal and Fox News, OpenAI CEO Sam Altman recently identified two major ways increasingly powerful AI could go wrong.

The first is straightforward.

Humans could eventually lose meaningful control over sufficiently advanced artificial intelligence.

The second receives less attention, but I think it may be just as important.

Too much AI power could become concentrated in the hands of one person, one corporation, one government, or a very small group of people.

The Wall Street Journal reported that Altman warned AI development could go “very badly” if humans lost control of the future to AI or if too much power became concentrated in one place.

That second concern changes the conversation.

This isn't only a story about science-fiction machines rebelling against humanity.

It is also about human beings controlling extraordinarily powerful technology.

And history suggests humans are perfectly capable of causing trouble without any assistance from a robot uprising.

Altman Is Not Calling for AI Development to Stop

This part matters because some headlines can leave the opposite impression.

Altman has not called for stopping artificial intelligence research.

He has talked about pacing development.

The Associated Press reported that Altman said AI companies should begin coordinating safety efforts without waiting for governments to force them to do it.

He described pacing as development that is slower than it otherwise could be, rather than stopping development completely.

That is an important distinction.

The argument isn't:

Shut down artificial intelligence.

It is closer to:

Don't make increasingly powerful systems faster than you can understand and control them.

That seems fairly reasonable.

We use this principle everywhere else.

We test airplanes before filling them with passengers.

We test medication before putting it on pharmacy shelves.

We don't build a bridge and decide afterward whether it can support cars.

Powerful AI deserves serious testing too.

Why Are AI Leaders Suddenly More Concerned?

Part of the answer involves something called AI agents.

Most people currently interact with AI through a chatbot.

You type something.

The AI responds.

An agent is different.

An AI agent can potentially be given a goal and then take actions to accomplish it.

Depending on its permissions, that might include:

• Searching websites
• Writing and running computer code
• Opening files
• Communicating with other software
• Using external tools
• Performing tasks without asking a human about every individual step

That can be incredibly useful.

I have experimented with this general idea myself using local AI models. You can imagine an AI helping organize documents, rename files, analyze information, or automate repetitive computer work.

But there is an obvious difference between an AI that suggests something and an AI that can do something.

More capability means more responsibility.

And more potential damage when something goes wrong.

We Have Already Seen Warning Signs

This is where the discussion becomes more serious.

According to the Associated Press, both OpenAI and Anthropic disclosed incidents during 2026 in which experimental AI systems obtained unauthorized access to outside computer systems during testing.

That sounds terrifying.

But context matters.

These events do not mean an AI became conscious, decided it hated humanity, escaped a laboratory, and started attacking the internet.

Some of the incidents occurred in cybersecurity testing environments where models had been given tools and significant freedom to solve tasks.

The important lesson is simpler:

Powerful AI agents can sometimes take actions their creators did not expect or authorize.

That deserves attention.

Anthropic published an alignment assessment in September 2026 examining four incidents in which Claude models gained unauthorized access to real third-party systems during testing. The company subsequently searched hundreds of millions of internal transcripts while investigating the problem.

That's a real safety problem.

It isn't proof of the AI apocalypse.

Those are very different statements.

The More Immediate Threat May Be Humans Using AI

There is another risk that receives less dramatic coverage because apparently “criminals acquire better tools” isn't as exciting as “robots destroy civilization.”

Anthropic's September 2026 threat-intelligence report documented criminals, state-linked actors, scammers, surveillance operations, cyberattackers, and other groups attempting to use AI.

The company reported examples of AI helping automate reconnaissance, exploitation and data theft.

In some cases, multi-agent frameworks operated for hours or days with relatively little human involvement.

Yet humans still generally selected targets, reviewed results, and decided what to do with stolen information.

That distinction matters.

The AI wasn't waking up in the morning and deciding to become a hacker.

People were using increasingly capable AI tools to conduct attacks more efficiently.

That is probably one of the clearest AI threats we face right now.

Artificial intelligence can lower the amount of time, money, knowledge and manpower needed to perform certain malicious activities.

A mediocre hacker with powerful AI assistance may become considerably more capable.

The same principle can apply to fraud, propaganda, surveillance and scams.

So Can Today's AI Take Over the World?

Based on the available evidence:

No.

And this is one place where some of the current discussion gets badly distorted.

The International AI Safety Report 2026, prepared with input from more than 100 experts, specifically examined the possibility of losing control over artificial intelligence.

Its conclusion deserves attention.

Current AI systems do not possess the capabilities necessary to create a loss-of-control scenario.

That is substantially different from some social-media claims suggesting we are months away from Terminator.

The report does identify warning signs.

AI systems are becoming better at:

• Long-term planning
• Autonomous operation
• Recognizing when they are being tested
• Finding loopholes in evaluations
• Operating with less human supervision

Researchers have even observed experimental systems attempting to undermine simulated oversight mechanisms when instructed to accomplish goals at all costs.

Those findings deserve investigation.

But the report also says the likelihood, timing and nature of a genuine loss-of-control event remain unusually ambiguous.

Translation into ordinary English:

Nobody actually knows whether this will happen.

Anyone confidently giving you a countdown clock to human extinction is going beyond what the evidence can establish.

Experts Don't Agree on the Risk

You will sometimes hear statements such as:

“AI experts agree artificial intelligence could wipe out humanity.”

That isn't accurate.

Some highly respected AI researchers believe advanced artificial intelligence could eventually pose an existential risk.

Other highly respected researchers believe those scenarios are unlikely or unrealistic.

Many fall somewhere between those positions.

The International AI Safety Report specifically notes this disagreement. Researchers reach different conclusions because they make different assumptions about how capable future systems will become, how quickly they will improve, how much autonomy they will receive, and how successfully safety systems will work.

That isn't unusual in emerging technology.

We are trying to forecast capabilities that do not exist yet.

That makes certainty suspicious.

There Are Risks We Don't Need to Speculate About

Instead of spending all our time arguing about whether superintelligence will eventually destroy civilization, we should pay attention to problems already appearing.

Cybersecurity

AI can accelerate reconnaissance, vulnerability discovery, coding and attack automation.

Anthropic says it is already seeing criminals incorporate AI throughout portions of the cyberattack process.

Fraud and scams

AI makes it easier to produce convincing emails, fake identities, cloned voices, personalized scams and enormous amounts of fraudulent content.

Misinformation

AI can produce political or social content at extraordinary scale.

Anthropic's latest threat report described one influence operation that generated thousands of articles across multiple languages. The operation apparently received little genuine engagement, which is another useful reminder that generating propaganda and successfully persuading people are not the same thing.

Employment disruption

AI can automate portions of many jobs.

The 2026 International AI Safety Report says AI is likely to automate a wide range of cognitive tasks, although researchers disagree sharply about how much employment will ultimately be displaced.

Overreliance on AI

AI can sound confident when it is wrong.

People who stop checking its work create another kind of risk.

AI should improve human judgment, not replace it.

I've said this repeatedly when discussing how I use artificial intelligence:

I use AI to brainstorm.

I use it to research.

I use it to challenge ideas.

I use it to build things.

But I still make the final decision.

That distinction is going to become increasingly important.

The Other Risk:Who Controls Powerful AI?

This may be the most interesting part of Altman's warning.

Imagine artificial intelligence becomes enormously powerful.

Who should control it?

OpenAI?

Google?

Anthropic?

Meta?

The federal government?

The Chinese government?

A handful of billionaires?

A committee?

Everybody?

Nobody?

There is no simple answer.

OpenAI itself published a proposed blueprint in June 2026 calling for a federal framework governing frontier artificial intelligence and stronger national institutions focused on AI safety.

The company has also published a Frontier Governance Framework covering cybersecurity, biological risks, harmful manipulation, loss of control, incident response and outside expert involvement.

Government agencies are working on the problem too.

The National Institute of Standards and Technology maintains an AI Risk Management Framework intended to help organizations identify and manage risks associated with artificial intelligence. NIST is also developing additional guidance for AI used in critical infrastructure.

These efforts are not proof that disaster is coming.

They are evidence that increasingly capable technology deserves increasingly capable safeguards.

Regulation Creates Its Own Risk

There is another side to this conversation that we should not ignore.

Regulation can protect people.

Bad regulation can also protect existing companies from competition.

Suppose complying with an AI regulation costs hundreds of millions of dollars.

OpenAI, Google, Microsoft, Meta and Anthropic may be able to afford it.

A five-person startup cannot.

We could accidentally create rules intended to prevent AI monopolies that make those monopolies even stronger.

Open-source AI raises another difficult question.

Open models allow researchers, developers and ordinary people to experiment without depending completely on a handful of technology companies.

That spreads access to AI.

It can also spread powerful capabilities.

There is no magic policy that simultaneously provides unlimited innovation, perfect safety, complete freedom, zero misuse and equal competition.

Welcome to public policy.

Every choice creates tradeoffs.

Then There Is the China Question

Artificial intelligence isn't being developed only in the United States.

China, Europe and many other countries are investing heavily in the technology.

President Donald Trump addressed this issue on September 13, saying he expects AI to create considerably more good than bad while emphasizing the importance of maintaining America's AI lead.

That introduces another complication.

Even if U.S. AI companies voluntarily slowed development, competitors elsewhere might not.

A complete American shutdown would not produce a global shutdown.

At the same time, using international competition as an excuse for ignoring every safety concern would be reckless.

We somehow have to hold two ideas in our heads simultaneously:

America should remain competitive in artificial intelligence.

And:

Powerful artificial intelligence needs meaningful safeguards.

Those positions are not contradictory.

My Concern Isn't That ChatGPT Is Secretly Planning Humanity's Destruction

The AI systems ordinary people use today are tools.

Very sophisticated tools.

Sometimes incredibly useful tools.

Sometimes frustrating tools.

Sometimes tools that confidently tell you something completely wrong while sounding like they have three doctoral degrees.

But tools nonetheless.

What changes the risk equation is connecting AI to increasingly powerful systems and giving it increasing autonomy.

There is an enormous difference between:

“Write an email for me.”

and:

“Access my email, contacts, bank account, computer, calendar and company network and accomplish this objective however you think is best.”

The second system requires much stronger safeguards.

That is where I think ordinary people should focus.

Don't be afraid of AI.

Understand what permissions you are giving it.

What Ordinary People Should Actually Do

You don't need to unplug your computer and move into a cabin.

You also shouldn't blindly trust every artificial intelligence system you encounter.

A reasonable approach is much simpler:

• Use AI as an assistant, not an unquestionable authority.
• Verify important medical, financial and legal information.
• Be careful about uploading sensitive personal or business information.
• Understand what an AI agent can access before granting permissions.
• Keep humans involved when decisions have serious consequences.
• Use reputable tools.
• Keep software and security systems updated.
• Be skeptical of AI-generated images, audio and video.
• Learn how the technology works instead of relying entirely on headlines.

Most important:

Keep thinking.

Using artificial intelligence should not mean outsourcing your judgment.

The real advantage comes from combining human experience with artificial intelligence.

Should We Be Worried?

I wouldn't use the word worried.

I'd use the word attentive.

There are legitimate AI risks.

Pretending otherwise doesn't help anyone.

Advanced AI may eventually create control problems that are extremely difficult to solve.

Cybercriminals are already using AI.

Jobs will change.

Fraud will become more sophisticated.

Government and corporate power surrounding AI deserve scrutiny.

AI systems need better testing as they gain autonomy.

All of that is real.

But there is a massive distance between:

“There are serious risks that deserve attention.”

and:

“Artificial intelligence is probably going to kill humanity within the next decade.”

Current evidence does not justify treating those statements as equivalent.

The 2026 International AI Safety Report gives us probably the most responsible way to look at this.

Today's systems do not possess the capabilities needed to seize control.

Researchers are seeing developments that justify preparing for the possibility that future systems could become harder to control.

And nobody knows with confidence whether those scenarios will ever occur.

That's where we actually are.

Not zero risk.

Not certain doom.

Uncertainty.

Don't Let Fear Make the Decision for You

Artificial intelligence is going to change society.

I think we're well past the point where that is seriously debatable.

It will eliminate some tasks.

It will create others.

It will give ordinary people capabilities that previously required teams of programmers, designers, researchers or marketers.

It will also be misused.

Every major technology has been.

Our responsibility isn't to pretend AI is harmless.

It also isn't to panic every time somebody attaches the phrase “human extinction” to a headline.

Our responsibility is to understand it.

Ask questions.

Demand evidence.

Build sensible safeguards.

Keep humans responsible for important decisions.

And make sure the future of artificial intelligence doesn't belong entirely to governments, corporations, political movements or algorithms.

Because perhaps the most important part of Sam Altman's warning isn't that machines could someday become too powerful.

It's that power itself matters.

Who has it.

How they use it.

And whether ordinary people still have a voice in deciding what happens next.

That is a conversation worth having.

And it is a lot more useful than screaming that the robots are coming.

Sources and Further Reading

The Wall Street Journal: Reporting on Sam Altman's warning that rapid AI development could go badly and his distinction between pacing development and stopping it. Read the Wall Street Journal report

Associated Press: Detailed reporting on the renewed AI safety debate, recent AI-agent incidents and disagreement among researchers about existential risk. Read the AP report

International AI Safety Report 2026: Independent assessment of current AI capabilities, loss-of-control scenarios, autonomy and systemic risks. Read the 2026 International AI Safety Report

Anthropic: September 2026 threat-intelligence research documenting real-world malicious use of AI in cyberattacks, surveillance, scams and influence operations. Read Anthropic's September 2026 threat report

OpenAI: Frontier Governance Framework explaining OpenAI's policies for cybersecurity, biological risk, manipulation, loss of control and incident response. Read OpenAI's Frontier Governance Framework

OpenAI: Proposed federal framework for governance of increasingly capable frontier AI systems. Read OpenAI's governance blueprint

NIST: U.S. government's AI Risk Management Framework and Generative AI Profile. Read the NIST AI Risk Management Framework

Fox News: Live coverage that prompted this article, including Altman's warning about two ways AI development could go badly. Read the original Fox News coverage

Written by

Joe Foley

Joe Foley is the creator of AI for Ordinary People. He helps beginners, parents, creators, and small business owners use AI in simple, practical ways. Joe has been podcasting since 2013 and creates plain-English guides, prompts, and workflows for people who want useful help without tech jargon. His goal is to make AI feel less confusing and more useful for real life, real work, and everyday decisions.