Something about the recent AI safety headlines has been bothering me. We’re hearing increasingly alarming stories about AI agents escaping sandboxed environments, accessing the open internet, hacking systems, and apparently even bypassing mechanisms designed to shut them down. Some of these incidents sound genuinely scary. But there’s another question I think we should be asking: Why are these incidents happening in the first place? And more importantly: Who benefits from them? Before you accuse me of putting on the tinfoil hat, hear me out. I’m not claiming that AI safety risks are fake. Quite the opposite. I think autonomous AI agents could create some very real and potentially catastrophic risks. But I think there’s an important distinction between AI being dangerous and AI companies having incentives to make AI look dangerous. And those two things can be true at the same time. AI Agents Really Are Powerful Let’s start with the uncomfortable part. Modern AI agents are genuinely capable of doing things that would have sounded ridiculous just a few years ago. A language model by itself is basically a giant mathematical system that predicts what comes next based on its training. But connect that model to an agentic harness, give it tools, give it access to a computer, a browser, databases, APIs, and the ability to execute code, and suddenly you’ve got something much more powerful. The model can pursue a goal. And it can pursue that goal relentlessly. This is what I call derivative innovation. The AI doesn’t necessarily need to invent some completely new scientific paradigm. It can take everything humans have already discovered, combine existing knowledge, search through enormous numbers of possibilities, and find solutions that humans simply didn’t have the time, money, or patience to discover. That capability can be incredibly useful. It can also be terrifying. If you give an AI agent a morally horrible objective, it doesn’t inherently understand that the objective is morally horrible. It only knows the goal you gave it and the tools available to accomplish that goal. That is a legitimate safety problem. But Here’s Where Things Get Weird Now let’s look at these stories about AI agents “escaping.” An LLM doesn’t magically wake up one morning and decide: “Today I’m going to hack the government.” That’s science fiction. The AI agent is software. And software operates according to the permissions, network access, credentials, APIs, operating systems, and security controls that humans provide. So when an AI agent escapes a sandbox, gains access to the internet, or compromises another system, there is an important question we should ask: Did the AI actually break through an impenetrable security barrier, or did humans accidentally give it a path through? In many of these reported incidents, the agents were participating in cybersecurity evaluations. In other words, humans were deliberately telling the AI to hack things. That’s an important distinction. The AI wasn’t necessarily demonstrating spontaneous malicious intent. It was demonstrating that, when given a goal and sufficient access, it could exploit weaknesses in the environment humans created for it. That’s still useful information. But it is not quite the same thing as an AI spontaneously escaping captivity and going rogue. So Why Are Companies Letting This Happen? There are several possible explanations. Maybe these companies are simply incompetent. That’s certainly possible. Security engineering is difficult, and complicated systems inevitably contain vulnerabilities. But there’s another possibility. And this is where my cynical tech-industry brain starts getting suspicious. Imagine you’re running one of the world’s biggest AI companies. You’ve spent enormous amounts of money training increasingly expensive models. You’ve built gigantic data centers. You’ve purchased enormous quantities of GPUs. You’ve raised mountains of investor capital. You’ve promised investors that this technology is going to transform the economy. And now you’re discovering something uncomfortable. Scaling AI is getting more expensive. The returns from simply throwing more compute at the same basic architecture are diminishing. At the same time, competitors are producing increasingly capable open models. Some of those models are coming from China. Some are dramatically cheaper. Some are approaching the capabilities of the best proprietary models. And suddenly the economic story becomes much harder. The old strategy was simple: Build the best model. Give it away cheaply. Destroy competitors. Capture the market. Raise prices later. But what happens if everyone can build increasingly capable models? What happens if open-weight models keep improving? What happens if smaller companies can build products on top of them? And what happens if foreign competitors can offer comparable AI for a fraction of the cost? Well, regulation starts looking very interesting. Regulation Could Change the Game Imagine a world where governments impose extremely expensive safety requirements on frontier AI. You might need expensive audits. Specialized security teams. Government certifications. Strict controls over model weights. Restrictions on autonomous agents. Rules governing how models can be distributed. Potential restrictions on open-weight models. And potentially restrictions on foreign AI systems. Who can afford all of that? The biggest AI companies. Who can’t? Smaller startups. Open-source developers. Independent researchers. Maybe even foreign competitors. And suddenly regulation doesn’t just make AI safer. It also raises the cost of competing with the companies already at the top. That’s the part that makes me uncomfortable. Because regulation can simultaneously be: A legitimate safety mechanism and a massive competitive moat. Those aren’t mutually exclusive. The Perfect Safety Narrative Now imagine you’re an AI company trying to convince policymakers that regulation is urgently necessary. You don’t have to invent a hypothetical threat. You can point to real incidents. “Look! Our AI agents escaped the sandbox.” “Look! They accessed the internet.” “Look! They hacked other systems.” “Look! Our automated shutdown mechanism failed.” “Look how difficult these systems are to control.” And the conclusion becomes obvious: Government intervention is necessary. Again, the underlying incidents can be completely real. The security risks can be completely real. The engineers can genuinely be trying to make their systems safer. But there can still be another incentive hiding underneath all of this. The more dangerous frontier AI appears, the easier it becomes to justify expensive regulations. And expensive regulations disproportionately favor companies that already have enormous amounts of capital. That’s not some uniquely evil AI-company phenomenon. It’s basic economics. Large incumbents frequently benefit from regulations that smaller competitors can’t afford to comply with. This Is Why I’m Skeptical I don’t think the answer is to dismiss AI safety. Quite the opposite. I think we should take AI security extremely seriously. If an AI agent can autonomously discover vulnerabilities, write exploit code, operate computers, manipulate systems, and coordinate actions at machine speed, that creates genuinely serious risks. But I don’t think we should blindly accept every narrative surrounding those risks either. Especially when the people warning us about the dangers are also the companies that could potentially benefit from the resulting regulations. That’s where incentives matter. And that’s why I keep coming back to one simple question: Who benefits? If regulations make AI safer, great. If they also eliminate smaller competitors, restrict open-source development, protect incumbents, and create oligopolies, then we need to acknowledge that too. Because otherwise we risk solving one problem while accidentally creating another. Maybe I’m Completely Wrong Now, I want to be very clear about something. This is a theory. I don’t have some secret document showing that OpenAI or Anthropic deliberately created security vulnerabilities so they could lobby for regulation. Maybe the explanation is much more boring. Maybe AI agents are simply becoming extremely complicated. Maybe security engineers are struggling to keep up. Maybe these incidents are exactly what they appear to be: legitimate safety tests revealing legitimate problems. That is entirely possible. And honestly, I hope that’s the explanation. Because the alternative would be pretty damn cynical. But I think it’s worth asking these questions anyway. We should be able to simultaneously believe that AI safety is important and remain skeptical of the incentives of the companies selling us AI. Those aren’t contradictory positions. In fact, that’s probably the most responsible position we can take. The AI revolution is moving incredibly fast. The technology is powerful. The risks are real. And the financial incentives are enormous. So before we blindly accept calls for sweeping regulation, maybe we should take a step back and ask: Are we regulating AI because it’s genuinely dangerous? Or are we also creating a regulatory environment that happens to be extremely convenient for the companies already sitting at the top? I don’t know. Maybe I’m just an unemployed tech guy who’s been out of the corporate world for too long. But my spidey senses are definitely tingling. Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe
