I’ve always loved studying history even though I didn’t always do great in those classes because they were more about memorizing dates and names. What I found interesting was the relationships. It’s probably one reason I love to read alternate universe sci-fi like Eric Flint’s 1632 (free download from Baen Publishing). For me, it inspires research into what really happened.

This editorial was originally going to be about the Robot Operating System and its importance to physical AI software, but the latest news flurry about controlling escaping AI models changed that. Robots, drones, and physical AI are also in the news as AI is now an integral part of these devices. However, the problem is not limited to these.

Anyway, back to the title. The Hugging Face Incident, as it’s known, had a bunch of AI agents working together to hack the Hugging Face website. The site is a platform for sharing “machine learning models, datasets, and applications.” It could have been any site.

The big difference between hackers of old and rogue AI agents is a combination of speed and access. Hackers needed to create or modify scripts or applications to get their job done. Replicating the person was hard, although replicating the number of active scripts and applications was easy. Access to the target was usually restricted.

On the other hand, AI agents can easily replicate themselves and have access to lots of resources, usually ones given to them. The ones they should not have access to are of concern. Those are also ones that AI models and agents have proven good at finding and exploiting.

Guardrails… We Don’t Need No Stink’n Guardrails

How many times have you heard a programmer who uses C say that they don’t need guardrails? These programmers are smarter than the average programmer and never make mistakes. It’s all those other programmers who caused the code to have bugs that can be taken advantage of by AI agents. All of this C coding is done in the name of efficiency as if developing secure and efficient code was hard.

We’ve Heard It All Before

The push for companies building a monopoly and exploiting resources isn’t new. Just think about the history of everything from railroads and trolley cars, oil to minerals. The need for guardrails, like worrying about the environment or people, was important then. But the impact on them was typically overlooked or, more likely, actively hidden.

The advances in these areas and the advantages, especially monetary ones, tend to be obvious. However, the longer-term impact can be significant, from famine to global warming. Even now, naysayers promoting these industries continue to point to the advantages and ignore or disparage those who point out that guardrails might be a good idea.

The call for guardrails and oversight is often used to cut off competition as well. The problem in these cases is that the guards and oversight often becomes advantageous to the current crop of companies and a challenge for newcomers.

In the end, it comes down to money and liability. The money tends to go to the larger, established companies with minimal liability. Lose a couple million social security numbers and a company just needs to pay for a year’s worth of another company’s theft protection services rather than the life of the information or person affected.

When it comes to AI agents, the liability is likely to have gone the way of subscription contracts with little or no liability on the agent’s side and a lot on your side. You’ll be lucky if you only get hit with mandated arbitration with a financial limit.

The AI Security Nightmare

Using a firewall on a network is comparatively easy with known traffic flowing through one place. The same is true for a building. Limiting the number of entry points makes it simpler to secure.

The problem with the AI agents these days is that the connections are vast and usually unknown to users and often developers. By design, AI agents can have lots of doors. Even now, security for agent-to-agent (A2A) and model context protocol (MCP) interfaces is still in its infancy.

Unfortunately, what’s behind those interfaces can lead to even more connections. Unlike conventional programming application programming interfaces (APIs), the A2A and MCP are much more flexible in what’s being sent and received to get the job done.

The other security aspect is that AI agents are inherently programmable and that’s often very easy for anyone, including another AI agent. Once through a virtual door, an AI agent can take advantage of not only the data resources, but compute resources as well.

Everyone is a Beta Tester

Finally, the fun begins with bait-and-switch tactics for AI adoption. The end goal is to get everyone to pay for AI-based services because they’re “so much better.” It’s why there are paid tiers for many chatbots, but most people use the “free versions,” acting as testers and trainers for future versions. Mix in a little targeted advertising and your beta testing days are paid for. Remember, you signed a contract to use this stuff.

I haven’t tied together all of the threads, but hopefully you’ve been able to follow them. I’m hoping real people will benefit from the ideas.

This rambling missive was done without AI assistance, as is the editorial for the latest Electronic Design Weekly that’s managed by real editors.

P.S. By the way, I asked a chatbot about the title of my article. It’s still thinking about it. “This may take a while…”

P.P.S. If you can find our article comments, please leave one. Comment About the Article is down below somewhere. Good luck finding it.