Hugging Face unveils ProvenanceGuard for source-aware AI agent verification

Hugging Face has released ProvenanceGuard, a verification method that matches AI agent claims to their sources and blocks incorrect attributions. The pipeline runs local MiniLM and DeBERTa NLI models on 2 GB of RAM and 2 CPU cores, or about $0.0001 per call via cloud APIs.

• ProvenanceGuard verifies not just facts but whether cited sources are correct
• It uses local MiniLM and DeBERTa NLI models
• Runs on 2 GB of RAM and 2 CPU cores
• Cloud API costs about $0.0001 per call, adding 200–300 ms latency