Theseus Guard
The World's First Stateful AI Governance Platform
5 followers
The World's First Stateful AI Governance Platform
5 followers
Theseus Guard is a stateful AI governance platform built for the Agentic Internet. Unlike stateless guardrails that evaluate interactions in isolation, it tracks agent identity and behavior across multi-turn sessions, nested agents, and lifecycle changes. It combines real-time runtime enforcement, behavioral drift detection, and tamper-evident cryptographic provenance—so enterprises can govern autonomous AI continuously, enforce policies at inference time, and prove what happened.
#Update - Sep 26'
We were finally able to complete the first end-to-end tests of our fine-tuned TG-Sentinel models. We subjected them to 2 group tests, including cross-sector, OWASP, and Red Team datasets.
These are not mere simple guardrail models; they do not simply classify content as "safe" or "unsafe." Instead, they evaluate inputs based on regulations and standards applicable in the US, EU, Turkey, and five other nations, as well as sector-specific attack vectors and AI risk factors. They then provide a response that explains the specific laws, rules, or procedures guiding the action taken. Within our "Decision Evidence" module, the system logs the specific layers responsible for the models' decisions—thereby ensuring explainability—and records these choices into an append-only governance log layer.
The results are as follows: These represent the native performance of the models, without any RAG, MCP, etc.
(Yes, we don't have a cool AI-generated infographic; what we have is a simple Excel spreadsheet.) 🙂
Special thanks to NVIDIA and Google for the GPUs. 🫶
US and EU-based companies will be able to test TG-Sentinel models (excluding the Ultra version) via API. You will be able to test them using the Open-Source CT-Toolkit (v0.33.7+) and its supported integrations—LangChain, DeepAgents, or MCP Server.
We are currently developing our proprietary commercial AI model, the TG-Sentinel:G1 35B-A4B, based on a hybrid BTX MoE architecture. It could well be the world's first AI governance model that is familiar with the regulations and standards of countries across three continents and has learned about attack vectors and AI-related risks. We have not yet encountered an AI model developed in this manner.
This is NOT a fine-tuned Qwen 35B-A3B or other model! We are continuing discussions with the BSC EU AI Factory (supercomputing center) and other institutions regarding training processes. We are aware of the requirements concerning costs, time, and resources.
With Theseus Guard, companies and institutions will be able to confidently transition their AI solutions from the pilot phase to production. We continue to work for the safety of us all!
Recent Developments:
We have been accepted into Microsoft, Amazon, and NVIDIA startup programs. (Evaluation processes for the other 2 programs are ongoing.)
Soon you will be able to use @Theseus Guard on your own VPC from Azure, AWS, and GCP cloud marketplaces.
We are continuing our work on Confidential Computing nodes.
We have revamped our website for a better experience.
We are nearing completion of integration work to allow you to use our TG-Sentinel:Fast stateful guardrail model for free within our open-source solution CT-Toolkit (https://hakandamar.github.io/ct-toolkit).
We have updated the reporting section for OWASP TOP 10 for LLM risks in Theseus Guard Enterprise. You can now see the compliance status of all relevant regulations.
We continue to strengthen the stateful governance layer and improve the reasoning routing feature for the TG-Sentinel:Max model. (Zero-Data-Memorization feature will only be available with TG-Sentinel:Ultra version)
We continue to work hard. ;)
AI agents need more than content filters. They need runtime governance. 💪
We are proud to announce that Theseus Guard, Inc. has filed a U.S. provisional patent application (64/149,945) covering a stateful, tenant-specific governance architecture for AI agent executions. 🥳
Our approach is designed to resolve governance policies at decision time, selectively evaluate risk, authorize tools and external actions, enforce bounded outcomes, and preserve auditable provenance for every governed decision.
This milestone strengthens our intellectual property foundation as we work toward making enterprise AI more accountable, secure, and operationally trustworthy.
From AI experimentation to governed execution. The journey has begun. 🚀
We tested Theseus Guard Enterprise according to the OWASP Top 10 LLM 2026 document published 5 days ago. The initial results are as expected. ;) The LLM model you use with Theseus Guard doesn't matter. TG protects against even the most current risks.
Meet the Zero-Trust AI Data Plane:
For organizations using cloud or remote LLM APIs, Zero-Trust AI Data Plane prevents sensitive PII and/or proprietary company information from reaching external models while preserving the task-relevant context needed for high-quality responses.
More than a PII filter: Zero-Trust AI Data Plane is a policy-aware runtime mediation layer that governs AI interactions before they reach external LLMs, protecting sensitive data while preserving the context required for useful model responses.
Secure AI communication without sacrificing usefulness!