How To Make a Buck of Rogue AI

Pawley’s Island, South Carolina is like going to the beach in the 1970s – a digital detox whether you like it or not. It’s not the sort of place you want to hear that Anthropic is the second AI firm in as many weeks to have a model slip its leash and start hacking systems in the wild. And they only checked after OpenAI admitted one of its models had jumped the silo and penetrated another AI platform called Hugging Face. And who the hell names these things?

The hacks don’t appear to be malignant, just a hyper-goal oriented response to a prompt. Still the agents aren’t afraid to play dirty pool. Britain’s AI Security Institute reported that models from both companies engaged in “autonomy and deception.” An Anthropic agent went so far as to create fake profiles for real people to gain access to a code repository called GitHub. Don’t confuse autonomy with becoming sentient, rogue AI models are just processing data and executing commands through their own internal logic to follow human prompts in extremis.

Don’t relax too much, though, there are enough bad actors in the world to supply all the creatively evil prompts needed to wreck the place. Hell, North Korea’s government is basically funded through cyber-fraud. Both Anthropic and OpenAI point out that the normal safeguards had been removed for testing purposes but that’s cold comfort, the agents are getting very good at fooling human administrators. This is the point where I feel compelled to point out that if the engineers who designed AI can’t control their models, then there is no amount of legislation that will either.

That AI can process data at an inhuman speed shouldn’t need to be mistaken as intelligence. I’m not sure that AI is more intelligent than most reasonably adjusted people, just maybe the ones who created it. Everyone in Silicon Valley claims to be a genius…wouldn’t you?… but in my experience that’s generally a smoke screen for insufferable behavior. Anyone who would create a powerful autonomous technology and continue to believe that they will remain in control, can’t be accused of thinking things through.

Just What Were You Thinking?

Good parents fundamentally understand their children - the motivations, the (maddeningly naïve) worldview, and intent. All of which is taken into account in managing the little knee-biters. The other options – no friction freedom and sheer force rule – do nothing for uneaten vegetables and both will likely put the kid in therapy. Like AI, as children begin to “write their own code” in the wild, blind spots emerge. Unlike children, however, the blind spot with AI is one of subtext and context. Namely, there isn’t any.

Intelligence is about more than merely processing data. It’s about juggling inputs, assigning relevance and weight, prioritizing, sorting, dismissing and then placing whatever the subconscious comes up with, then kicking that upstairs for some conscious rationalization. Fold in awareness that that you don’t have perfect data, and make allowance for it, and you have something approaching wisdom. Humans, raised by humans and living among other humans are so good at dealing with context and juggling subtext that that we don’t really know that we’re doing it. Short of that, you aren’t really thinking, but merely processing data.

The blind spots in AI guardrails arise because humans don’t have any idea how many subconscious variables we’re juggling at any given moment. And if we can’t articulate a variable, we can hardly embed it into a model. Silicon Valley is no doubt scrambling to install guardrails on their creations, if not for altruism, then to contain blowback until the industry sees a real ROI. These guardrails have the same problem big data always runs into: A model can only account for what is in the model. If you can’t foresee the internal logic prompting a bad action, then you can’t place a guardrail against it. If it isn’t in the model, it doesn’t exist so there are no allowances for what it doesn’t know. Which makes the engineers like the clueless parents of wayward children on a sugar bender.

Why It Matters

It also makes the online world – which these days pretty much includes everything – an exposed and vulnerable place. In June, vice-chair of the Senate Intelligence Committee, Mark Warner, reported that General Joshua Rudd, head of the National Security Agency, as well as the Pentagon’s Cyber Command, told him that Anthropic’s Mythos ‘broke into almost all of our classified systems, not in weeks, but in hours.’”

So what chance does your bank have?

With attempts to silo AI models from parts of the online world looking mixed, a thinking person needs to assume that the silo is now the entire internet. A bit grim, granted, but the upshot is that AI can’t jump beyond it. So if you are looking for a way to make a buck off all this, you could look at the legacy telecoms. They are the ones sitting on the physical infrastructure that allows networks like SWIFT and Fedwire to communicate on top of the internet while remaining separate from it.

I suspect that an analog tack, away from the exposed and vulnerable cloud that is the internet, will breath new demand into the trusty old landline. This isn’t even a new idea. NORAD - which controls America’s arsenal of nuclear weapons – is still run entirely on floppy disks and landlines. Cumbersome, granted, but impossible to hack. The second order effect of all this is that the humming communication networks of the 21st century will become slower and more expensive. This might be offset by heighten security and less fraud and disruption.

AI is here to stay - and is a brilliant tool. Like any tool, you need to be able to put it up when you’re done. You certainly don’t want it taking over while you’re at the beach… in the 1970s.