TECNALIA develops VAIBreaker to test model security

TECNALIA has developed VAIBreaker, a tool that assesses the robustness of AI models by testing their responses to threats that could disrupt their operation. It subjects models to controlled attacks that simulate adverse scenarios, identifying vulnerabilities before deployment in operational environments.

Beyond detecting weaknesses, VAIBreaker measures how attacks affect model performance, including reduced accuracy, poorer prediction quality and diminished trust in results. It translates those findings into objective metrics and practical recommendations intended to help developers, integrators and organisations strengthen models and make informed decisions about their use.

The evaluation process can be repeated throughout a model’s life cycle to support continuous improvement. VAIBreaker also provides documented evidence to help organisations meet requirements from companies, users and regulatory bodies. Its focus is on understanding model limitations as AI takes on increasingly critical functions in sectors including industry, healthcare, mobility and digital services.