The framework covers cybersecurity, information safety, content material, human-AI interplay, and bias. Its cybersecurity class consists of malicious code era, malicious software calls, extraction of the mannequin’s security insurance policies, extreme token era, and habits that diverges from the supposed system objective.
The paper attracts on present NIST and ISO risk-management requirements and is the primary in a deliberate collection on managing dangers throughout open-weight AI deployments, the businesses stated in a weblog put up.
Mannequin adjustments can weaken safeguards
Open-weight fashions might be fine-tuned, merged with different fashions, modified, or repackaged for various makes use of. However these adjustments can weaken a mannequin’s security protections even when it continues to carry out its supposed activity effectively, Microsoft and Nvidia stated.
