MITHRIL SECURITY was a Paris company that built open-source tools to make AI more private and more verifiable. Its products let AI providers serve models on confidential data without being able to see that data, and prove how a model was trained.
It worked for AI providers and the organisations that use them, where sensitive data and regulated models meet. Every guarantee rested on secure hardware, so a user could check a cryptographic proof rather than take the provider’s word.
At A Glance
| Vendor | MITHRIL SECURITY |
|---|---|
| Category | Confidential AI and AI model provenance |
| Deployment | Open-source tools running on TPMs and secure enclaves |
| Customers | Future of Life Institute, Avian Digital Forensics, Auxlia, Scille |
| Founded | 2021, Paris |
| Market classification | Agentic AI Identity |
What MITHRIL SECURITY Does
Two trust problems hold back the use of AI on sensitive data. Users have to hand their data to the AI provider, and they have no proof of how the model they are using was trained. The company set out to solve both, giving AI providers a secure supply chain with provenance traceability, model protection and data confidentiality.
The approach was to make trust checkable. Instead of policies and contracts, the user verifies the environment and the model directly before relying on them. The tools are open source and have been audited by independent security labs.
Products
BlindLlama
BlindLlama is a zero-trust AI API for running open-source LLMs on confidential data, for organisations that cannot let a model provider see their inputs. An open-source Python SDK on the client checks the hardened environment before sending anything, and open-source server enclaves serve the model with leakage channels such as SSH, logs and network access removed. It was released as an alpha.
AICert
AICert is for model builders who need to prove what went into a model, for example that copyrighted or biased data was not used in training. It creates a certificate that binds the model’s hash to a hash of the training procedure. Training code is supplied as a Docker image and run on secure hardware, and certificates can be stored on a public ledger. It is published under the Apache 2.0 licence, and was still in development with a call for beta testers.
A third tool, AIGovTool, provided trusted monitoring of AI and was developed in collaboration with the Future of Life Institute.
How It Works
| Component | What it does |
|---|---|
| Secure hardware | The tools rely on TPMs or secure enclaves, which attest that a specific software stack is loaded, from the BIOS through the operating system to the application. BlindLlama uses secure hardware that cloud providers make available. |
| Attested TLS | For BlindLlama, a reverse proxy sits in front of an attesting launcher that loads the hardened container and produces the attestation proof. The client then checks a TPM-measured hash of the server certificate before sending any data. |
| Certificates | AICert runs training on provisioned secure hardware and shares the resulting certificate so anyone can verify the model’s provenance. |
Company Background
| Founded | At Station F in Paris, France. |
|---|---|
| Leadership | Daniel Huynh, CEO. |
| Funding | 1.2 million euros raised in June 2023. The company also received an OpenAI Cybersecurity Grant in January 2024. |
| Acquisition | Acquired by H company in the second half of 2024. Part of the team, including Daniel Huynh, joined H. |
Note: This overview is compiled from MITHRIL SECURITY’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.