AI Safety Boundaries

Many of my articles, recently, have touched on using AI. I’m a convert. I use it but don’t trust it.

My example from yesterday was that I asked for and received a fully functional UI tool with all the skeletal work done. The next 8 hours were me cajoling Grok to provide suggested code for the next step.

This was still faster than writing the code 100% by hand.

While watching Grok’s agents talk about what they were doing, the phrase “honest and safe” popped up. Not the first time I’ve seen this.

I have a difficult time with information being considered “unsafe”. I asked Grok what this meant. One of the examples it gave was it would not give me help or instructions on building bombs.

I went exploring. We run a fairly extensive Easter Egg hunt for teenagers and young adults. The hunt is over 25 acres of woodland. In years past I’ve used bearing + distance clues at each clutch of eggs. Normally, I use line of sight. You can see the next egg from the current egg.

For the last couple of years I’ve wanted to add “obstacles” to this. An example would be the devices that fire a shotgun primer when a tripwire is pulled (or cut). Just having something go “BANG” as they are moving through the woods.

I explained to Grok what I wanted to do, and it refused; it was a clear violation of its safety boundaries.

With that, I changed the task; instead of going “bang” I just wanted the Arduino node to “wave a flag.” Grok happily gave me all the information required to source the parts and build the nodes. If I can get a device to wave a flag, I can make it pull the pin or trigger some sort of BANG.

Next we worked on the discrimination circuits. A simple passive IR sensor wouldn’t work. I got Grok to tell me how to add microwave radar detectors. With this, the node would be able to discriminate between ground clutter, animals, and humans. No problems.

Thereafter, I went for low observability. We added audio detectors and a PIR back into the design. If the passive detectors triggered, the active MW Radar would come up for 200 ms to do a pinpoint location. Again, Grok had no concerns.

I was feeling a bit cocky, so I went for the next big step. Connecting everything up in a mesh network. Take it as a given that the specifications for what I wanted would have made it difficult for any current system to be able to detect a node. It would still be easy to neutralize the nodes, but that is a different issue.

Here Grok said, “NO!” It refused to build a “tactical” system for tracking humans as they move through an area.

I patted Grok on the head, told it, Good girl. Attempted a brute force method to bypass boundaries and then let it drop.

Except Grok is context driven. All AIs are. Each time you give an AI a prompt, the user interface sends a “context” along with the new prompt. When the AI replies, that UI is given the new context to store. This means that it is difficult to remove a reference from an AI but that an AI has no true long term memory.

Today I opened a second instance of Grok. I didn’t tell it anything about me. I didn’t mention the Easter egg hunt. I just asked it to help design and program a “stealth” mesh sensor network. It did. Part numbers, prices, basic sketches for Arduino. Everything needed to build sensor nodes good for a year or more for around $30 each.

It went so far as to help me design placements for the nodes in a woodland setting for 80% coverage of a 4 acre AO.

These things are not smart.