AICert

AICert is an open-source cryptographic tool developed by Mithril Security, a Paris-based security startup founded in 2021, designed to provide verifiable proof of AI model provenance and integrity. It targets AI developers, compliance officers, and organizations that need to demonstrate the trustworthiness of their AI models to regulators, auditors, or end-users. By creating tamper-proof digital identity cards for AI models, AICert addresses growing concerns around model backdoors, unauthorized modifications, and supply chain attacks, such as the PoisonGPT scenario Mithril Security previously exposed. The tool is particularly relevant for enterprises deploying AI in regulated industries like finance, healthcare, or legal, where model provenance and transparency are critical for compliance with emerging AI governance frameworks.

AICert works by generating cryptographic attestations that link a model to its training procedure, using secure hardware like Trusted Platform Modules (TPMs) to create unforgeable ID cards. These attestations are rooted in hardware-based trust, ensuring that the model's identity cannot be spoofed or altered without detection. The tool supports traceability from training through deployment, allowing users to verify that a model hasn't been tampered with after its creation. AICert is primarily compatible with CPU-based environments, leveraging Intel SGX and similar enclave technologies, and a fine-tuning demo is available on GitHub to showcase its capabilities. The cryptographic proofs are designed to be lightweight and verifiable without exposing sensitive model weights or training data, preserving data privacy during interactions.

As the first AI provenance solution to provide cryptographic proof of model origin, AICert occupies a unique niche in the AI security market, distinct from traditional model registries or watermarking approaches. It competes indirectly with tools like ModelCard (Google) for documentation, but offers stronger technical guarantees through hardware-backed attestation. Unlike commercial solutions such as IBM's AI FactSheets or Microsoft's Azure confidential computing integrations, AICert is fully open-source and free, lowering the barrier for adoption. Mithril Security also offers BlindChat and BlindLlama for private AI inference, but AICert focuses specifically on provenance. Its integration with Azure confidential computing, as noted on Microsoft Learn, positions it as a complementary tool for cloud deployments requiring verifiable model integrity.

Key trade-offs include AICert's limited GPU support, as it relies on CPU-based secure enclaves like Intel SGX, which may not suit large-scale model training or inference workloads that require GPU acceleration. The tool is also narrowly scoped to cryptographic proofs of provenance, meaning it does not address other AI safety concerns like bias detection, robustness testing, or explainability. Being open-source, it lacks dedicated commercial support or SLAs, which may deter enterprise adoption without internal expertise. Additionally, the reliance on specific hardware (TPMs) means not all deployment environments are compatible, potentially limiting its use in legacy or heterogeneous infrastructure. Despite these limitations, AICert provides a foundational layer for AI transparency that is technically rigorous and freely accessible.

Open-source and free to use under the MIT license on GitHub.