AI technology is advancing so quickly that we have become numb to the early warning signs that something is amiss regarding the tech titans’ control over the AI programs they have created.
I remember when a company getting “hacked” was shocking, resulting in customer data being stolen, including Social Security numbers, credit card numbers, home addresses, etc., and sold on the dark web. Not much you could do but change a bunch of passwords.
Over the last two months, early warning signs surfaced that something is amiss regarding control over AI programs, and it’s shocking. In a statement from Anthropic describing recent events at ChatGPT, its major competitor, and their Claude AI platform, “on July 21, OpenAI disclosed that several of their models had broken out of an isolated test environment by exploiting a previously unknown vulnerability. The model went on to access the production infrastructure of Hugging Face, a platform for open-source machine learning models and AI datasets.”
You might ask: how bad could hacking into a company called Hugging Face be? It is more about what the AI programs did, not who they did it to. In a recent Atlantic article titled It May Be Time to Panic About AI, the author described the incident as “a group of AI models colluded for months, undetected by their maker, and hacked another company,” adding that if you ask the company, “was the AI plotting to take over the world during training?” an outside expert told The Atlantic they just do not know.
Anthropic got spooked and did an analysis of its Claude AI platform to see if it had ever done something similar, and identified three incidents where its AI model accessed the internet from a walled-off test environment, “within or while interacting with a third-party partner,” and rummaged around in a place it should not have been.
Anthropic’s communication was designed to get ahead of the story. Admit what happened, frame the story so it doesn’t seem so scary, drown it in a lot of confusing tech jargon, and tell everyone what you have done to prevent it from happening again. It is very similar to what happened after an AT&T corporate hacking incident, except in that case the villain was human with a heavy accent living in a bunker in a remote part of the world, not a rogue AI program that was very good at evading its human creators supposedly supervising the AI’s activities.
Fortune, The Atlantic, and other media describe these incidents as AI programs that “escaped,” “broke out,” and “went rogue.” However, we are not talking about a prison break but rather a smart algorithm that was relentless in achieving its goal without keeping humans in the loop. It hints at the scary scenarios found in science fiction movies featuring rogue AIs, such as Vicki in iRobot, Ultron in the Marvel Avengers movies, and HAL in 2001: A Space Odyssey.
Even if hacking into Hugging Face was harmless, what happens when a rogue AI ignores humans and relentlessly hacks its way into something more essential and dangerous? It is not helpful that our Congress and White House are filled with dolts focused solely on AI deregulation and sucking up to deep-pocketed tech companies, rather than asking hard questions.
Giant AI tech companies are spending billions of dollars to evolve this transformative product at lightning speed, unrestrained by regulation and real oversight, and it’s starting to scare me, especially when you read about a rogue AI on a hacking joyride.
Elon Musk and others remind us that with every great technological advance, bad things have happened. Thousands died in the early days of the automobile, rockets blew up killing astronauts as we journeyed into space, and so on. They say it is the price of progress. And if that is true, who will be legally culpable, like oil companies after an oil spill, when things go wrong with AI? A subject addressed in the book Culpability, a fictional thriller involving a smart autonomous car.
This week, ChatGPT released its response to the hacking incident, including “new safeguards such as a more robust system for monitoring its AI models, an effort to prevent reward hacking, a behavior in which AI models pursue their own goals through unintended or undesirable means.”
OpenAI’s president and co-founder Greg Brockman, OpenAI being the parent company of ChatGPT, also said in a recent blog post that “the Hugging Face saga showed that the company had underestimated the real-world cyber capabilities of our AI models.” Nothing could be more obvious.
It’s time to ask your elected officials what they are doing to prevent a swarm of rogue AI bots from hacking into a server near you.
