#servicenow

ServiceNow’s MosaicLeaks benchmark targets a quiet failure mode in deep research agents: private facts leaking through external queries. Training only for task success raised leakage from 34.0% to 51.7%, while PA-DR cut it to 9.9%.

Anthropic announced on January 28, 2026 that ServiceNow selected Claude as its default model for AI agent development. ServiceNow cited up to 95% productivity gains in some workflows and reported large-scale AI request volumes.