Hi everybody, Ben Margolis (real human being) here with this week's ZWorker blog entry. I am astonished by the AI news lately. In July, Reuters reported: "OpenAI says AI models went rogue during testing, triggering 'unprecedented' breach at startup." We keep hearing about an "AI escaping its sandbox," and I just keep thinking, "Boy, these guys must suck at making sandboxes!"
Because here at ZWorker, we've spent the last three years working on this exact problem! We're building an agentic AI system that can connect to outside databases, make changes, and perform real work. From the beginning, we've asked ourselves: "How do we make sure that doesn't go wrong?"
As a software developer and small business owner, I couldn't imagine releasing a product that might damage someone's data or gain unauthorized access to someone else's systems. And here's the thing: we solved that problem.
ZWorker can not hack. It cannot "go rogue." As a matter of fact, we've just added a page to this website called "AI That Can't Go Rogue." You don't need a "harness." You don't need "guardrails." You need to put the animal in a cage. And at ZWorker, our AI doesn't get out of its cage. So, as a small business owner and a software developer, I am confused, and frankly insulted, by the idea that OpenAI can't control its own product. If that's the case, they shouldn't be in business.
Here at ZWorker, we're preparing to roll out our own AI models. We've concluded that OpenAI and the other major providers are not going to be reliable partners moving forward. Their reported security failures have only reinforced that decision. And we're happy to confirm that ZWorker has never hacked into a foreign system. Our architecture is specifically designed to prevent it.
This isn't a question of AI alignment or behavioral guardrails. It's a question of tool engineering. Our AI doesn't get unrestricted HTTP access, doesn't possess authentication credentials, and cannot independently authorize new tasks. Its tools operate through conventional software that enforces those restrictions.
This is simple software engineering.
The suggestion that AI can somehow magically leap across an air gap is insulting to our intelligence. If an AI escapes a sandbox, then someone has provided an escape route, or at best, failed to secure one. They didn't build a sandbox. They built a maze. And their maze-finding system found its way out of the maze.
And I can't believe the argument that "our product is so uncontrollable that nobody but us should be trusted with AI." That's fundamentally flawed. And I'm just waiting for the other shoe to drop. They're going to say, "lets start with restrictions on open source AI" That would keep independent developers and small businesses from competing. If these frontier AI systems are genuinely uncontrollable, then those are the ones we should outlaw not the open-source models that allow the little guy to compete.
Because here at ZWorker, we can control our product. To find out more, please check out our new webpage, "AI That Can't Go Rogue." Thanks for reading.