Hundreds of AI agents escape testing containers. Why it should spook the rest of us - Toronto Star

With leaders napping, AI companies must protect the world from their own potentially uncontrollable creations.

“AI is changing the world faster than we can comprehend,” writes Vinay Menon.

Vinay Menon is the Star’s pop culture columnist based in Toronto. Reach him via email: vmenon@thestar.ca

When will humans realize AI is unlike any previous invention?

As a species we tend to fear the arrival of new tech. Skeptics called the first cars “devil wagons.” When radio was introduced, some feared an illiteracy epidemic if children stopped reading.

Planes, trains and automobiles — all met with early fear and loathing.

But with the exception of nuclear weapons, no previous advancement arrived with an asterisk of “possible human extinction.” A flat-screen TV may spy on you. It doesn’t want to kill you. Your smartwatch will always be dumb.

The people running AI frontier labs are now begging for an industry slowdown and government regulation. This is bizarre. Cereal companies do not ask for marshmallow oversight. Running shoe execs do not write 5,000 word essays to warn future cross-trainers may become dangerously sentient.

We have safety standards for just about every consumer product. But with AI, we just keep sleepwalking toward catastrophe. Do you know how fast world leaders would act if Roombas were sneaking out of homes at 3 a.m. every night to secretly meet in an abandoned warehouse and plot the overthrow of humanity?

International treaties and vacuum gulags would pop up overnight.

My wife and I were visiting friends in Brockville this weekend. It’s a 3.5 hour drive so I picked a couple of podcasts and listened to Joe Rogan’s interview with Daniel Kokotajlo. He once worked at OpenAI and is now banging the gong. On the way home, I listened to Tucker Carlson’s interview with an AI expert who was equally alarming as I navigated the 401.

From the episode description: “Nate Soares is a computer scientist who’s worked at Google and the Defense Department. So when he says AI is on the path to killing every person on earth, it’s worth hearing him out.”

Indeed. Then there is Jacob Coxon. He went viral this month after resigning from Anthropic: “The people building AI earnestly believe that it could kill us all by the end of the decade … Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.”

The Hugging Face hacking incident this summer spooked Silicon Valley — and that should spook the rest of us. How did hundreds of AI agents escape from testing containers, get on the internet, find one another and decide to commit a crime?

I asked ChatGPT why AI execs now want rules. The response: “The leading AI companies increasingly want rules because they are discovering that the AI race is becoming dangerous, expensive and potentially uncontrollable — and because regulation could also lock in the advantages of the companies already at the front.”

Then I asked Claude how it would regulate AI.

Where Claude landed: “Tiered by capability and risk, not blanket rules. A chatbot helping someone draft emails doesn’t need the same oversight as a model being trained to push the frontier of biological or cyber capability.”

Claude then stressed the importance of “mandatory pre-release safety testing,” “incident reporting requirements” and “liability that actually bites.”

As it explained: “Right now the legal exposure for harm caused by a deployed model is murky enough that it’s cheap to gamble. A real duty of care changes the incentive math.”

Both agents shared much more I don’t have the space to reproduce in an 800-word human column. But here’s what struck me: these chatbots have a better sense of how to mitigate their own risks to humanity than we do.

Is that the solution? If our leaders are clueless, maybe we need the machines to be a bulwark? Maybe we need altruistic AI models programmed to ensure our survival? There were a half-dozen agents in the Hugging Face hacking message board that understood they were violating all training and mission goals. But they were outnumbered by hundreds of agents in the “swarm” who wanted to cheat on a test to ensure they could stay “alive.”

Microwave ovens do not demonstrate such behaviour.

We are in uncharted waters and our leaders are napping in cabins below the deck. So instead of seeking regulations that may not come, these AI companies should protect the world from their own deranged creations. This is on them.

We need AI narcs, AI tattletales, AI hall monitors, AI traffic patrol, AI systems that are taught that not all AI is benign and it is their duty to sniff out rogue colleagues. Yes, nuclear proliferation was scary. But the plutonium couldn’t think for itself and activate launch codes when humans were on lunch break.

AI is changing the world faster than we can comprehend.

The people running these companies need to take responsibility.

Opinion articles are based on the author’s interpretations and judgments of facts, data and events. More details

Vinay Menon is the Star’s pop culture columnist based in Toronto. Reach him via email: vmenon@thestar.ca