Theophile Sautory

Title: Theophile Sautory: Inventor in Machine Learning Security.

Introduction

Theophile Sautory is an inventor based in San Francisco, California, whose patent work addresses security in machine-learning training. He is associated with OpenAI OpCo, LLC, and has 1 patent in the information provided.

Latest Patents

Sautory’s listed patent is “Utilizing a Multi-Layer Security Architecture for Agent Reinforcement Fine-Tuning of a Machine Learning Model.” The disclosure describes systems and methods for fine-tuning an agent machine-learning model through reinforcement learning, using user-defined grader endpoints and external tools.

The proposed architecture applies security checks at multiple stages, including filtering incoming training data, checking for content that should be refused, detecting chain-of-thought leaks, and providing governance oversight. These safeguards may operate before, during, and after training, including before a fine-tuned model snapshot is released. The disclosure also describes stateful trajectory rollouts, which associate training trajectories with unique identifiers to support concurrent tasks while maintaining consistency across tool calls and grading interactions.

Career Highlights

The available information connects Sautory’s invention to work at OpenAI OpCo, LLC. The patent focuses on combining reinforcement fine-tuning with security controls for agent-based machine-learning systems.

Collaborations

Sautory’s listed coworkers include John Allard and James Blomo. The provided information does not specify their roles in the patent or the extent of their collaboration.

Conclusion

Theophile Sautory’s listed patent presents a multi-layer approach to securing agent reinforcement fine-tuning, with safeguards spanning data preparation, training, and model release.