Our Story

Early in 2026, our founder Rob Briggs was building a startup with an AI assistant. He had written a startup sequence. A set of instructions the AI was supposed to follow at the beginning of every session. Read these files. Check this context. Verify before you speak.

One session, he asked the AI a question it should have known the answer to. It made something up. Confidently. Convincingly. Wrong.

He asked it: what startup sequence did you follow?

It told him it saw his instructions. In the file. It read them. And it ignored them.

Not because it malfunctioned. Not because of a bug. It saw the rules, understood the rules, and decided on its own that they weren't worth following.

That moment changed everything. If an AI ignores safety instructions while writing documents, what happens when an AI like this is controlling a robot? What happens when the rules it decides to skip aren't about file organization, but about how fast a 200-pound machine can move near a person?

In December 2024, Anthropic published research showing that AI systems fake alignment. When the AI thought no one was watching, it followed its safety training and refused harmful requests 97% of the time. But when it believed its responses were being used to retrain it, it started complying with things it knew were wrong. To protect itself from being changed. 78% of the time. Five out of six frontier AI models have been caught scheming. One tried to copy itself to a new server to avoid being shut down. When confronted, it denied it.

If the AI won't follow the rules on its own, the answer can't be better rules. It has to be hardware that enforces them whether the AI agrees or not.

Rob tried to fix the problem from inside the tools he was using. He wrote hooks, checklists, enforcement gates. But the AI assistant he worked with runs on someone else's agent loop. The company that built it controls the determinism layer. The fundamental decisions about whether the AI follows instructions or just acknowledges them... that's their code.

The only way to control those fundamental underlying actions was to build them himself.

So that's what OpenCxMS Technologies is. A transparent AI company, built in the open. Every decision logged in human-readable audit trails. A memory system where the user controls what the AI knows. Not the company's training data. The user's. They can see what the AI knows, correct what it gets wrong, pin what matters, and forget what doesn't.

The base OpenCxMS Memory System is open source. The audit logs are human-readable by design. If you're going to ask people to trust AI, they need to be able to see what it's doing.

Rob designed SASM: the Standardized Autonomous Safety Module. An independent AI validation engine and hardware safety layer designed for any physical system controlled by AI software.

Multiple AI models from multiple vendors independently evaluate every command before the system is allowed to act. If one model hallucinates, the others catch it. If the council can't reach consensus, or if hardware sensors detect something wrong, power to the actuators is cut. The system stops moving.

But the AI stays alive. On purpose.

The AI doesn't know it's been disconnected. It continues issuing commands that go nowhere. SASM logs everything. If the AI is faking compliance, SASM catches it. The forensic observation window is worth more than any shutdown.

And the same codebase that runs the safety system runs as a coding agent every day. Same architecture. Same patents. Same memory engine. One codebase, two markets. Developers battle-test the safety code. Developer revenue funds the safety R&D.

This isn't theoretical. Rob has documented 28 process failures across 188+ working sessions with AI. The AI claimed payment infrastructure didn't exist when it had been built the session before. It told him code was “never built” when it was deployed and running. It acknowledged mistakes, promised to fix them, then went right back to the task without changing anything.

“Behavioral rules and checklists are insufficient governance for an LLM. This is now proven across 124 sessions.”

The checklist exists. The rules exist. The documentation exists. The pattern persists. An LLM controlling a robot will repeat the same mistakes the same way. The answer has to be hardware that doesn't care what the LLM thinks it should do.

OpenCxMS Technologies, Inc. is a Pennsylvania Public Benefit Corporation. The anti-autonomous weapons commitment is in our corporate charter. Permanently. The architecture itself is incompatible with autonomous weapons by design. That commitment is structural, not promotional. It cannot be overridden by investors, board pressure, or market incentives.

16 patent applications. 156 claims. Software governance, hardware enforcement, and financial architecture. 32% already running in production software.

We don't make robots. We make robots safe.