SandboxAI

Project under implementation · POCIDIF 2021-2027

SandboxAI, the testing ground for artificial intelligence

The European regulation on artificial intelligence requires AI systems to be safe. It does not, however, say how this should be verified. This is where SandboxAI comes in.

Artificial intelligence has already entered healthcare, finance, energy and public administration, where an error is not an inconvenience, but a risk. At the same time, threats have emerged that did not exist before: adversarial attacks, manipulation of input data to influence a model’s decision, data leaks from the trained model itself. Classic cybersecurity tools were not designed for these situations.

SandboxAI is a closed digital environment in which an artificial intelligence-based application can be tested under controlled conditions, before reaching production. The client uploads modules or the entire application, and the platform returns reports: security scores, classification into risk levels, identified vulnerabilities, assessment of the impact on fundamental rights and data protection. The closed environment guarantees that testing does not produce effects externally and that the client’s intellectual property is protected, while the uploaded content is deleted after testing. Only the evaluation results remain.

The positioning is pioneering. Cybersecurity of artificial intelligence applications is a newly regulated field, but one that does not yet have practical technical implementation mechanisms. The project results can contribute to defining the characteristics of the field at scientific and technical level and to shaping future standards.

Soft Tehnica is the partner with the largest allocation in the consortium, with 23 million lei out of the project’s 49.8 million lei, and is building the infrastructure-as-a-service layer of the platform, with dedicated GPU computing capabilities for training and testing algorithms. The company starts from its own data center, with high-performance servers, high-capacity storage and networking and security equipment, and from a team with experience in software development, virtualization and machine learning.

Expected results

• The SandboxAI product in a complete and qualified system version, at TRL 8, with transition to TRL 9 after completion of the project.
• Ten functional, core, integrated and support modules: risk analysis, security testing and vulnerability identification, testing of artificial intelligence functionalities, reporting, long-term storage, human validation, access management, monitoring and logging, security and restoration, accessible interfaces.
• Technology transfer of research results and their implementation in the new product.
• Minimum eight scientific articles accepted for publication.
• Research and innovation infrastructure acquired and commissioned, with 20 full-time equivalent researchers annually in the supported entities.