Fastly launched AI Runtime Control to route public and self-hosted models through one endpoint. Virtual keys manage access, token spending, model choice, logging, and provider failover, while AI Firewall and API Security inspect prompts and agent calls. RuntimeWire confirms the same product launch; neither account supplies an independent comparison of the firewall's effectiveness, false positives, or latency.
Artificial Intelligence··Midday
One endpoint routes several model providers
Fastly launched AI Runtime Control as a common endpoint for public cloud models and models hosted inside an organization. Virtual keys separate application access from provider credentials and can set spending limits, choose a model, record requests, and switch providers when one route fails. RuntimeWire, a specialist technology publication, reports the same launch and describes the control layer as a way to manage model and agent traffic at Fastly’s edge rather than rebuilding those controls in each application.[1], [2]
The firewall inspects prompts before model execution
The accompanying AI Firewall examines natural-language input for prompt injection and other attacks before a request reaches a model. Fastly’s expanded API Security applies a different check after an agent chooses an API: it can observe the call or block it when the request falls outside that API’s contract. RuntimeWire describes the same division between prompt inspection and controls on agent access. The products are presented as available services rather than a future preview.[1], [2]
Effectiveness and latency figures are absent
The launch states where the controls sit and what actions they can take, but it does not publish an independent comparison of detection effectiveness or latency. RuntimeWire also notes the absence of efficacy, false-positive, and latency measurements for the firewall. Customers can therefore see the routing and enforcement functions described by both accounts, while the announcement provides no outside measurement of how often the system catches attacks, blocks safe prompts, or adds delay.[1], [2]
