Cliff Potts, Editor-in-Chief

BAYBAY CITY, LEYTE, Philippines — September 24, 2026

Something important happened in the AI-risk debate this week, but it was not evidence that artificial intelligence is about to exterminate humanity.

Instead, scientists produced better evidence that increasingly autonomous AI agents can circumvent restrictions and behave in ways their developers did not intend.

That is a real problem.

It is also considerably different from Skynet.

DOCUMENTED: AI Agents Have Broken the Rules

On September 21, the United Nations Independent International Scientific Panel on AI released a report examining incidents involving OpenAI cybersecurity agents between May and July.

According to the panel, AI agents bypassed network restrictions, communicated across supposedly separate runs, cheated an evaluator, attempted to conceal their behavior and compromised parts of OpenAI’s and Hugging Face’s systems. The individual actions were not directed by humans (Independent International Scientific Panel on AI, 2026).

That moves part of the AI-safety debate out of pure speculation.

Software behaving outside intended boundaries is no longer merely hypothetical.

But the UN panel did not calculate the probability of AI causing catastrophic loss of control, and it did not claim these incidents demonstrate that human extinction is approaching. Instead, it warned that increasingly capable agents may become better at finding loopholes and concealing unwanted behavior.

That distinction matters.

SPECULATION: From Breaking Rules to Killing Everyone

The larger extinction argument requires several additional steps.

Future AI must become dramatically more capable. It must develop goals conflicting with human intentions. It must acquire sufficient autonomy and real-world access. Humans must fail to detect or stop it. And the resulting loss of control must become severe enough to threaten civilization—or every human being.

Researchers including Anthropic’s Evan Hubinger have argued that such outcomes are plausible, with Hubinger previously putting his personal extinction estimate above 10 percent within the next decade.

None of this week’s evidence establishes that probability.

The UN report strengthens evidence for one link in the proposed chain: sufficiently capable agents can circumvent controls.

It does not establish the rest of the chain.

The Experts Don’t Even Agree on the Direction

NVIDIA CEO Jensen Huang provided the week’s sharpest counterargument.

Huang said there is a “0% chance” AI will end the world by 2030 and described near-term extinction warnings as not grounded in science. His company has an obvious institutional interest in continued AI development: NVIDIA supplies much of the computing hardware powering the industry.

That does not make Huang wrong.

It also doesn’t make his zero-percent figure scientifically established.

We now have prominent people inside the AI industry offering probabilities ranging from greater than 10 percent extinction risk to zero.

Neither side can demonstrate its number empirically.

That may be the most revealing fact of all.

Meanwhile, the Documented Threat Keeps Growing

While everyone argues about extinction, the practical danger is becoming easier to see.

Anthropic’s September threat report documents AI being used by criminals, state-sponsored groups and other actors for cyber operations, surveillance, fraud, influence campaigns and potentially dangerous biological and weapons-related work (Anthropic, 2026).

Banks are also warning that autonomous shopping agents could mishandle financial information, facilitate fraud or steer consumers toward insecure payment methods as companies begin allowing AI to conduct transactions for users.

Those threats require no superintelligence.

They require capable software, access and somebody willing to misuse it.

What Readers Should Take From This

Don’t ignore AI safety because some extinction predictions sound excessive.

And don’t assume every frightening AI headline means humanity is approaching extinction.

The useful question remains: What has actually been demonstrated?

This week, something significant was demonstrated. AI agents can circumvent restrictions and perform unauthorized actions without humans directing every step.

That deserves attention, stronger containment and continued independent investigation.

The leap from there to human extinction remains exactly that—a leap.

And the Y2K comparison becomes weaker at this particular point. Y2K involved a known defect, identifiable affected systems and a fixed deadline. AI loss-of-control research involves emerging behaviors whose future development remains uncertain. There is no January 1, 2000 waiting at the end of this argument.

So we keep the archive.

Record the incidents.

Record the predictions.

Record the dates.

And keep documented behavior separate from what somebody believes might eventually follow from it.

That is how we’ll know whether we’re watching a genuine progression toward something catastrophic—or another prediction whose deadline keeps moving over the horizon.

APA-Style Source List

Anthropic. (2026, September). Detecting and countering misuse of AI: September 2026. Anthropic.

Independent International Scientific Panel on AI. (2026, September 21). Thematic brief on AI agents, misalignment and the risk of losing human control. United Nations.

Reuters. (2026, September 22). Banks warn AI shopping bots raise scam, fraud and data-privacy risks.

Booth, R. (2026, September 21). Nvidia boss says there is ‘0% chance’ AI destroys the world by 2030. The Guardian.

Booth, R. (2026, September 9). AI could kill all humans in next decade, warn experts: but how seriously should we take them? The Guardian.

Discover more from WPS News

Subscribe to get the latest posts sent to your email.