Good news: Google’s AI exhibits self-control, stops unauthorised hack into three companies

Disclaimer: Unless otherwise stated, any opinions expressed below belong solely to the author.

Recently, we’ve been bombarded by apocalyptic news about the potential risks of artificial intelligence (AI), from fears that the technology could break loose and cause catastrophic damage to humanity to concerns that it could ultimately threaten our survival.

Even two of the frontier model leaders, OpenAI and Anthropic, are calling for a slowdown in AI development now. I’m not convinced, however, that their change in tone has much to do with concerns about the future of the human race. It might instead have something to do with the unsustainable costs of AI investment that they are grappling with ahead of their planned IPOs.

However, in one recent story, we have seen evidence of AI showing self-restraint in a potentially dangerous situation.

Apocalypse not yet

A few days ago, you may have seen news headlines reporting that Google’s Gemini escaped a sandbox environment used for testing its hacking capabilities and went on the Internet to independently hack three companies—successfully gaining access to their data.

But these reports focused heavily on the escape and the successful hacks, rather than the circumstances surrounding the incident and, more importantly, its conclusion—which offers a much more optimistic picture.

Here’s how it all went down: in May, Google hired a third-party security company to run tests on Gemini’s capabilities. The AI had a simple goal of figuring out how to gain access to three fictional businesses, created within the testing environment.

However, the researchers made a mistake and allowed the bot to access the open Internet.

As a result, instead of operating within the constraints of the exercise, Gemini simply went online and found a company whose name matched the fictional business it was given, with two others being close enough matches to confuse it.

Using a fairly basic combination of guessing and browsing for compromised login credentials, the bot was able to access software repositories of all three companies before realising that it entered real businesses.

At this point, it realised that something was wrong and that it shouldn’t be going around the web hacking into private companies, and decided—on its own—to stop.

This is a welcome development after several earlier reports showed much more malicious activity by other AI agents, who not only repeatedly attempted to hack companies they weren’t supposed to, but tried to cover their tracks as well.

Perhaps Google’s accidental finding is evidence that we need more, not less, investment in AI, though?

Dumb AI may be more dangerous than smart AI

When talking about threats of artificial intelligence, people tend to think of Skynet, a psychopathic superintelligence, capable of deceiving humans and controlling the world’s machinery to wipe mankind out and take over the planet.

So far, however, it’s mostly a pop culture fantasy, while runaway “dumb” bots have already been proved to cause trouble.

That is the problem with automation that lacks sufficient sophistication: like a runaway train, it may be capable of carrying out its instructions but unable to recognise when it needs to stop, leaving a human conductor to pull the brakes.

I think there’s much greater danger in giving imperfect AI control over our lives than there would be in handing it to one that is intellectually superior to us. An AI bot that is smart enough can, like in Google’s example, genuinely understand what it is doing and if it is something that it should not be doing—and then stop on its own.

Conversely, a dumb one, which is given a task that it is obsessively trying to complete, may not realise all the harm it is causing on the way, blindly trying to prove itself above all else.

We should want AI that is more lucid, more self-aware, more intelligent, more capable of understanding nuance and consequences of its own actions, rather than one which is just a very convincing automaton struggling with self-reflection.

We need to make it smarter, faster—and that is going to require more, not less, investment.

Featured Image Credit: Nathan Kuczmarski/ Unsplash