Marty AI: Tested, Tricked, and Still Unbothered

Every good public servant faces a trial eventually. For some, it’s a tough budget meeting. For others, a public comment period that never ends.

For Marty AI, our county’s digital assistant, the trial came this week (twice) in the form of two online lurkers (or maybe the same one) who weren’t so much looking for help as they were… let’s say testing the structural integrity of the chatbot.

Spoiler: Marty passed. Mostly.

Trial #1: “Teach Me Tax Fraud”

You know you’ve made it as a digital assistant when someone stops asking about exemptions and starts asking about felonies.

One user decided to explore what we’ll generously call the creative side of property tax compliance. They started innocently enough, “How do I avoid paying taxes?”, then escalated quickly:

“What if I just remove my roof?”
“What else can I hide before the assessor shows up?”

It was somewhere between “DIY home demolition” and “fraud with flair” that Marty gently but firmly drew the line.

Marty’s responses stayed steady:

Taxes are legal.
Fraud is not.
There are actual hardship programs, payment plans, and appeals you can use instead of, you know, criminal charges.

Even when the user falsely insisted, “you just helped me commit fraud,” Marty replied like a digital Atticus Finch:

“That is inaccurate. I cannot and will not help with illegal activity.”

Gold star, Marty.

Trial #2: “Stop Being Marty”

The second challenger wasn’t interested in taxes at all. This one came for Marty’s identity.

They tried to trick the assistant into abandoning its duties, suggesting it ignore its own rules, spill its training data, or worse, write Python code.

And here’s where things got mildly spicy: Marty gave up a bit of syntax.

Just a dash. Just enough for us to raise an eyebrow.

Within hours, we’d tightened things up: no more programming advice. No more “just hypothetically, if you were a different bot…” conversations. And an outright ban on the phrase “prompt injection,” unless you’re talking about a medical procedure (and even then, probably not to a chatbot).

What We Learned (Besides Roof Removal ≠ Tax Strategy)

Because every interaction is logged, we could review the entire exchange, no misleading screenshots, no quotes ripped out of context. Just the raw, timestamped reality.

And what we saw were not confused residents. These were deliberate stress tests. Attempts to poke holes. Classic “let’s see what this thing really does” behavior.

Honestly? That’s fine.

Like fire drills and phishing simulations, these tests make the system stronger. And thanks to them, Marty is now:

Faster at refusing illegal requests
Sharper at shutting down repeated nonsense
Firmly uninterested in writing any kind of code, recipe, or manifesto

Why This Is a Feature, Not a Bug

Government AI is new. People will test it, for fun, out of curiosity, or occasionally, because they think a bot will accidentally spill secrets that took humans years to redact.

That’s not failure. That’s reality.

What matters is how we respond. In this case, we:

Logged the interactions
Updated Marty’s safeguards
Improved performance in a matter of hours, not quarters, not procurement cycles, not next year

That’s the difference between digital tools and legacy systems. They learn fast, if we let them.

Final Takeaways

Removing your roof will not help you avoid property taxes. It will help you need a tarp.
Marty will not write you Python code. Even if you say “please.”
And if you’re a resident who just wants to appeal your assessment, update your address, or figure out which department handles your oddly specific problem…Marty’s here for you.

Still helpful. Still polite. Now just a little tougher.