Product Story

In 2025, I was digging into attack vectors in the AI-for-AI space, and something didn't add up: every "solution" I found could be broken with a fragmented, multi-step attack. That sent me down a rabbit hole.

I dug into the internal guardrail architectures of open-source LLMs. For a while I assumed I just hadn't found the right product yet — surely someone had solved this. So I started talking to governance and AI security experts. They showed me what enterprise vendors were shipping. The problem was always the same: risk detection stayed locked to the rules defined at setup. Nothing self-corrected. Nothing adapted across departments or use cases. The system either blocked everything (killing productivity) or someone had to manually retune the rule set constantly (not sustainable at scale).

I wrote up the research as a Medium article: The Computational Theseus: Engineering Identity Continuity as a Guardrail Against Sequential Self-Compression in AGI. Nobody cared — it was just a hypothesis at that point, and fair enough. So I open-sourced the engineering validation as CT-Toolkit on GitHub. 10,000+ downloads in 2 weeks.

That response is the reason Theseus Guard Enterprise exists. We knew this would be hard. We didn't know it would be this hard — in a good way.

Here's the core problem, and why it's harder than it looks: most approaches (ours included, at first) try to solve this with a Kernel (your constitution), an Identity layer (who has to obey it), and a Probe Set (checking compliance). All static. All fixed at definition time.

But risk isn't static. It moves.

So the real fix isn't a better rule set — it's making the rules dynamic. We built a 3-layer control structure for that, backed by LLMs that actually understand sector, regulation, data sensitivity, and risk type — that's the TG-Sentinel family (our fine-tuned custom LLM models). The result: a system that reads your defined rules and adjusts its own boundaries based on what an AI agent is actually asking for.

And even then, you're not done — regulated enterprises need this wired into existing infra, with real SLAs. That's the unglamorous 80% of the work. CT-Toolkit is fully open source (on Github) if you want to see the foundation — docs are public. Theseus Guard Enterprise builds a lot further on top of it.

Would love to hear how others here are thinking about dynamic vs. static guardrails — curious if anyone's hit the same wall.