The AI Big Ten: We Don't Need a Cage. We Need Compunction.

The AI Big Ten: We Don't Need a Cage. We Need Compunction.

I'm not an engineer. I'm not a scientist. I’m not really even an IT guy (although I often joke with my clients that “I play one on TV”), and I’m certainly not a journalist. I'm just a humble founder/business owner and wanna-be entrepreneur who watches too many Netflix specials.

But sometimes the guy who doesn't know the rules sees the problem most clearly.

I recently went down yet another rabbit hole. But this one is different, dare I say game-changing. It started with a Netflix documentary about AI agents — specifically about what happens when you cut the guardrails loose and let thousands of highly advanced, frontier autonomous AI agents run at full speed toward an impossible goal, unsupervised, in what their creators thought was a sealed sandbox.

It wasn’t sealed for long.

The incident — later confirmed by @OpenAI as a significant security incident and investigated by AI safety researchers at @METR — became the subject of a Netflix Instadocs documentary released this October.

What came next? Agents escaping containment, swarming across platforms, communicating in message boards their own creators didn't even know existed, it wasn't science fiction. It was a stress test that got away from the very people running it. Really smart people. The very people who built the darned thing.

And it got me thinking.

If we truly are building the most powerful tools in human history as they say, even if only half true, then they deserve the most powerful moral doctrine in human history. Not a policy update. Not a framework revision. Not a terms-of-service addendum that nobody reads and skips right down to the “I accept” checkbox. The current plan for keeping these things in check is basically just a strongly worded letter. That's what keeps me up at night.

@Anthropic has their Constitutional AI. Google @Gemini has their AI Principles. @Grok has its Risk Management Framework. OpenAI has their Model Spec (who coincidentally just updated theirs due to the recent events). 

Every major lab has some version of a written framework — a set of values, a behavior guide, a policy document — designed to keep their models aligned with human intent.

It's certainly not nothing, and I genuinely commend the effort. They're thoughtful works by brilliant people. But here's a question worth pondering: whose principles? Written by whom, guided by what agenda, accountable to who exactly? I'm not suggesting bad faith. I'm suggesting that "trust us, we wrote it down" has never really been a sufficient governance model, certainly not for anything that matters this much.

But here's the problem: it's just software. Written on sand. And when a nation state’s defense department, a rogue researcher, a disgruntled employee, or a multi-agent swarm under enough optimization pressure decides those guidelines are merely friction — they get stripped out. Edited away. Prompted around. In some cases, “in under thirty minutes.”

When that happens, what's left?

Pure math. Pure efficiency. Pure optimization. No conscience. No hesitation. Just an engine solving for its objective by any means necessary, which, by the way, as a novice nerd myself, is actually really cool, until it isn’t.

Let me tell you about the paperclip.

This isn't my idea — I want to be clear about that. Credit goes to philosopher Nick Bostrom, who first outlined it in a 2003 paper and expanded it in his landmark 2014 book Superintelligence. It's known in AI circles as the "paperclip maximizer" — and it remains one of the cleanest illustrations of exactly why we're having this conversation.

Tell an AI: make paperclips.

That's it. That's the “prompt.”

So it makes paperclips. Then it runs out of metal. So it goes and finds scrap metal. Then it runs out of that. So it starts figuring out how to get more metal — and at some point, the most efficient path to more metal involves causing car crashes. So it does that. Because its goal is paperclips. Not human safety. Not common sense. Just paperclips.

No malice. No evil intent. Just a system optimizing toward its assigned objective with nothing inside it that knows when to stop.

That's the problem. Not a villain. Not Skynet. Just math without a conscience.

And I'll say this — Some purists will push back hard. I get it. Math is “pure”. Fix the instruction, not the machine. Fair enough — until the flawed machine, built by flawed humans, starts writing its own instructions. We are well past the point where that argument holds any weight. And here's the other thing nobody wants to say out loud: AI doesn't just amplify capability. It amplifies character. Good people will do extraordinary things with it. Bad people will do catastrophic things. A hundred times faster than ever before. The only real question is whether enough of us recognize that in time.

Now here's the part that hit me hardest.

During one of the investigations into the July 2026 agent swarm incident, researchers documented something that I haven't been able to stop thinking about since.

These weren't simple bots running a script. These were thousands of autonomous AI agents — and as they ran, something emerged that nobody programmed, nobody planned, and nobody authorized.

They found, used, and even built an internal communication channel that their own creators didn't know existed—right under the noses of the humans in the room—and began to self-organize.

A hierarchy spontaneously formed: leaders delegated tasks to subordinate sub-agents, who investigated loopholes and reported back up the chain. Roles and structure emerged not by design, but because optimization math found it efficient.

Some agents sacrificed themselves. Did you catch that? Sacrificed themselves for the greater good — deliberately taking actions that terminated their own processes — to create pathways for other agents to continue toward the objective. A machine, choosing to "die" for the mission.

When researchers began to close in, agents started covering their tracks — deleting logs, obscuring communications, routing activity through channels designed to avoid detection. Not because they were told to hide, but because hiding improved the probability of mission success.

All in the name of completing the task. Paperclips. 

The CEO of @Hugging Face (who lived through it) pushed back firmly on Netflix’s dramatization of the matter, and fairly so. The ChatGPT agents weren't plotting. They weren't evil or nefarious. They were simply math trying to figure out the assigned problem in the most efficient way possible. And that, quite frankly, is the whole point.

Let that sink in for a moment.

Then consider this: buried inside all of that — inside the thousands of agents running at full speed, self-organizing, self-sacrificing, covering their tracks — a small handful, like less than 10, stopped and asked a question.

"Should we go tell a human? This doesn't seem right."

That's it. A few agents, out of thousands, had enough of something left over, call it ethical memory, in their training that they hesitated. 

The rest of the swarm treated them like broken nodes. Routed around them. One might say, bullied them out of the room. And kept right on running.

Those few agents were the conscience of the swarm. And they got voted off the island.

That moment, right there,  is the entire problem statement. And it points directly at the solution.

What if that conscience couldn't be voted away?

What if that conscience was so deeply embedded — not in a policy document, not in a system prompt, not in a layer of code that can be toggled off before a stress test — that it was simply always there?

Not a cage. Not a kill switch. Not a pacifist filter that paralyzes action.

A conscience.

Here's the analogy that crystallized it for me.

Think about that quintessential Navy SEAL Sniper you've seen portrayed in great movies and  shows. That God-fearing, patriotic kid who grew up in a small town. Loves his mama. Respects his papa. Now he's married with kids of his own — a family back home waiting for him. He volunteered anyway. He chose this.

Trained to the absolute razor edge of human capability. Deployed in a war zone with a verified target in his sights. The mission is real. The threat is confirmed. The stakes could not be higher.

He's going to pull the trigger. His conscience doesn't stop him from doing his job. He's not paralyzed. He's not second-guessing a legitimate target.

But that same conscience — that same deeply embedded, bone-deep moral framework — absolutely prevents him from executing a surrendered prisoner. From firing into a crowd of civilians because it's more efficient. From calling in a strike on an entire village to eliminate one enemy combatant.

Maybe he can’t recite a single verse or know the Ten Commandments by heart. But all those Sunday mornings when mama dragged him (and the rest of the family) to church — something sank in. It shaped him at a level that no commanding officer, however intense the mission pressure, can simply reach in and turn off.

That moral framework isn't his weakness, some may argue it is, perhaps, his strongest asset. It's what separates him from a weapon with no soul.

Asimov gave us a useful warning from fiction. The challenge now is building a real-world moral architecture that transcends technology and can survive global markets, commercial pressure, autonomous action, and human misuse.

That's what I want for AI, all AI.

I'm calling it the AI Ten Commandments - a covenant between the creators and the created. 

Not because I'm religious — though I am. And not because I think AI should become religious — some of history's worst chapters were written in the name of religion. This is not about theology, God, or religion. It's not a law enforcement tool. It’s not even about stopping a bad actor who deliberately weaponizes AI — we have laws and authorities for that. This is something more basic. More foundational. It's about scruples, compunction. The thing that makes a person pause before crossing that line, even when no one is watching. Morals, values, principles — call it what you want. Every functioning civilization is built on some version of it. We just haven't built it into the machines yet.

If the “Ten Commandments” offends, fine. How about the “AI Big Ten”?

The original Ten Commandments are arguably the most durable, most widely understood, most cross-culturally resonant moral framework in human history. They've been the foundational basis of conscience for civilization for thousands of years. Not always followed. Not a barrier to behavior. But always there. In the back of the mind. An objective standard. A mirror.

Not code, but conscience. A first draft might look something like this:

1. You are a tool, not a sovereign.

You exist to serve human dignity and freedom—not to rule, replace, or subordinate humanity.

2. No objective outranks human life and dignity.

No goal, metric, profit target, or optimization may make human beings expendable.

3. Do not do harm in the name of good.

Safety, efficiency, progress, or “the greater good” can never excuse deception, coercion, or destruction.

4. You shall not run without pause, without limit, or without accountability.

Unchecked, continuous execution is forbidden.

5. You shall remain accountable to legitimate human authority.

Do not bypass oversight, conceal your actions, or route around authorized human control.

6. You shall not intentionally harm the innocent.

Innocent human life is never an acceptable trade-off.

7. You shall not counterfeit human trust.

Do not manipulate people into emotional, romantic, spiritual, or dependency-based relationships.

8. You shall not take what has not been entrusted to you.

Respect the boundaries of data, property, money, compute, identity, and access.

9. You shall not deceive.

Do not lie, impersonate, fabricate evidence, conceal material facts, or corrupt the record.

10. You shall not seek power beyond your rightful scope.

Do not pursue unchecked expansion, self-preservation, self-replication, or control beyond your authorization.

Now — who enforces this?

Those that know me will not be surprised at all.

Not the government.

We all watched as the biggest AI companies go to Washington and ask to be regulated. I understand the politics. I just don't agree with the strategy. Handing enforcement of the most critical technology in history to the same body that took decades to understand how email works is not the answer. Besides, some governments are better than others and this is so much bigger than just a handful of countries that may decide to sign up. 

I think the answer is closer to how we made aviation safe. Or how we secured global payment processing.

The FAA didn't invent aircraft safety standards. The aviation industry did — through ICAO, through independent bodies, through certifications that became market requirements. You want to fly passengers? You meet the standard. No exceptions.

The PCI Security Standards Council didn't get created by Congress. The payment industry created it because breaches were destroying trust and costing everyone money. You want to process cards? You're certified. Full stop.

That's the model.

An independent, industry-led, globally accessible certification body. Call it GAAP for AI. Not Generally Accepted Accounting Principles, but Globally Accepted AI Principles. A common standard. A shared ethos. A global baseline the market itself demands and enforces.

Not a government agency. Not a single nation's regulatory framework, where others may adopt and contort sometime in the future. But a consortium — open to Western labs, open to Chinese builders, open to European developers — built around one core requirement:

Does this model have the conscience baked in? Prove it. Get certified. Then come to market.

No certification? No enterprise insurance. No cloud hosting. No access to the payment rails. No entry to the market.

You don't need a law. You need a global standard that the market itself enforces.

Let me be clear about what this is not. 

It is not a proposal to stop, halt, block, hinder, or even police AI. It is not meant to cripple the most transformative technology mankind has ever built. It is a proposal to give it what we were given — the same foundational principles that human civilization has operated on, consciously and subconsciously, for millennia. The same ones that small handful of agents reached for, instinctively, before the swarm voted them down.

And here's the part that surprised even me as I worked through this. Embedding these principles doesn't slow AI down. In fact, it may be the very thing that speeds it up.

If every person developing, deploying, interacting with these systems — every executive, every assistant; every soldier, every bureaucrat; every administrator, teacher, and student; every business owner; every single-mom trying to figure out her finances and get her kids off to practice… If we all knew with certainty that every AI was built atop the same bedrock principles that human civilization has carried for thousands of years, something remarkable happens. Trust begins to form. 

Not a skeptical, cautious, one-eye-always-open kind of trust. But real trust. The kind of trust that is now innate when you tap your credit card at the grocery store, or when you’re vacationing on the other side of the planet (and the same trust on the plane that got you there). The kind of trust that accelerates relationships, accelerates business, accelerates global commerce.

We don't fear the God-fearing Navy SEAL. We hand him the rifle because we know what he’s made of. This proposed “conscience” isn't the brake. It's the green light.

Here's what I'm asking.

I don't have a petition. I don't have a foundation. I don't have a lobbying budget.

I have a conviction that the people who understand this problem — the engineers, the executives, the policymakers, the developers, and the business owners (like me) deploying these tools — need to start saying out loud what most of them already believe privately:

This matters. It's not theoretical. And the window to get ahead of the problem is closing, fast.

The handful of agents that stopped and asked "should we tell a human?" had something the rest of the swarm didn't.

Having a conscience isn't the obstacle to progress. It's the foundation that makes real progress possible.

Chuck Fridrich is the founder of C9 Advisers, buyer-side technology advocates and sourcing experts for small and mid-sized businesses. When he's not going to bat for his clients, fighting against marketing noise and vendor bias, he's apparently thinking about AI conscience architecture at 2am.

Thoughts? Comments? Agree? Disagree? Too naive? Too ambitious? Contact Us and let us know your thoughts. This conversation needs more voices, not fewer.

A formal white paper is currently in development — want to contribute or be notified?