Microsoft and Nvidia urge provenance checks for open-weight AI
Original title: Microsoft, Nvidia push provenance checks as open-weight AI adoption accelerates
Their paper lists 30 risk types and says weight or config changes can weaken safeguards while task performance stays intact.
Microsoft and Nvidia published the first paper in a planned series telling enterprises to evaluate open-weight models through deployment. They identified 30 risk types across cybersecurity, data protection, content, human-AI interaction, and bias, and said fine-tuning, quantization, merging, and configuration changes can degrade safeguards without hurting routine task performance. The paper recommends cryptographic signatures to match a publisher’s artifact and separate tests of safety and capability after material changes. Gartner separately projects open-weight models will cover 50% of enterprise AI use cases by 2028, up from 20%.
Unconfirmed: Gartner’s 2028 share is a projection, and the unnamed Japanese-style gaps Gogia flags are his assessment, not measured incidents.
- Cyber risks named include malicious code, malicious tool calls, safety-policy extraction, and excessive token generation.
- Where signatures are unavailable, unverified integrity should stay an open risk.
- Gartner says open-weight inquiries rose more than 30% from Q1 to Q2 2026.
- Gartner also recommends AI gateways, adversarial testing, and an approved-model registry.
- 1.Verify model provenance with cryptographic signatures when available.
- 2.Test the exact configuration to be deployed and retest after material changes.
- 3.Route open-weight traffic through governed gateways with authentication and audit logging.
- Organizations
- GartnerGreyhound ResearchNIST
From the original article
3 min read at csoonline.com
Microsoft and Nvidia are urging enterprises to assess the security of open-weight AI models throughout deployment, warning that changes to model weights and configurations can weaken safeguards even when a model retains its ability to perform assigned tasks.
csoonline.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.