# Fastly puts model routing, AI budgets and prompt-injection controls into the edge request path Fastly has launched AI Runtime Control, AI Firewall and expanded API Security as production controls on its edge network, combining multi-provider model routing and spend limits with prompt-injection filtering and contract enforcement for agent traffic. The interesting part of Fastly’s AI launch is consolidation: model gateway economics, LLM security and agent-to-API authorization now sit in the same request path as the CDN/WAF infrastructure many applications already use. - Status: Active - Published: 2026-09-22T21:30:06+12:00 - Updated: 2026-09-22T21:30:06+12:00 - Categories: Artificial Intelligence, Cloud & Infrastructure, AI Agents, Edge & CDN, Inference & APIs - Tags: AI agents, API security, Fastly, model routing, prompt injection - Canonical HTML: https://beyondthe.news/dossiers/fastly-ai-runtime-control-firewall-agent-api-security ## What changed Fastly announced AI Runtime Control, AI Firewall and new API Security capabilities on September 21. Runtime Control provides a single endpoint across public and self-hosted model providers with virtual provider keys, token/spend visibility, rate limits, budget controls and failover. AI Firewall evaluates prompts in the edge request path to identify and block LLM-oriented attacks such as prompt injection. API Security can enforce API contracts against agentic and conventional traffic, observing or blocking requests that use unsupported operations or malformed payloads. Fastly says all three capabilities are now available. ## Why it matters Teams putting AI into production increasingly assemble separate products for model routing, cost governance, LLM security and agent authorization. Fastly is making those controls part of the edge infrastructure already carrying application traffic. That can reduce the number of gateways in a request path and make policy consistent across ordinary APIs, AI applications and autonomous agents. The trade-off is concentration: model access and AI security become more dependent on one edge control plane, and Fastly’s security/economic claims still need independent production evidence. ## Runtime Control behaves like an AI gateway with FinOps controls Applications can route model calls through one Fastly endpoint while retaining multiple public or self-hosted providers. Virtual keys keep underlying provider credentials out of applications, while token visibility, rate limiting, budget controls and failover move model economics and resilience into the routing layer. ## The firewall sits before the model AI Firewall evaluates prompts in the request path at the edge and is intended to block prompt-injection and related LLM attacks before they reach the target model. That differs from post-generation moderation or application-level checks because enforcement occurs on infrastructure already fronting the service. ## Agent traffic gets API-contract enforcement Fastly’s API Security expansion treats autonomous and agent-assisted requests as API clients that still have to obey service contracts. Operators can observe or block calls that target unsupported operations, send malformed requests or otherwise fall outside defined API behaviour. ## The architectural bet is one policy plane Fastly is combining AI routing and security with its existing bot management, WAF, DDoS and API tooling. For teams already on Fastly, that creates a path to govern human, application and agent traffic without adding another dedicated AI proxy. For teams that value provider independence, the same consolidation increases the blast radius of Fastly configuration or availability failures. ## Key details - Fastly announced the new capabilities September 21, 2026. - AI Runtime Control supports public and self-hosted model providers behind a single endpoint. - Runtime controls include virtual keys, token/spend visibility, rate limiting, budgets and failover. - AI Firewall is designed to detect and block prompt-injection-style attacks in the edge request path. - API Security can enforce contracts for agentic, agent-assisted and conventional API traffic. - Fastly says AI Runtime Control, AI Firewall and API Security are now available. - Fastly reports machine-generated traffic exceeded 50% of its network traffic in July and August; this is a vendor network statistic rather than an internet-wide measure. ## Builder takeaways - If you already front AI applications with Fastly, compare consolidating model routing and security there against operating a separate AI gateway plus LLM firewall. - Keep provider credentials behind virtual keys and test provider failover with real application semantics; successful HTTP fallback does not guarantee equivalent model behaviour. - Set explicit token/spend budgets per application or environment so runaway agents cannot consume an unrestricted shared model account. - Treat prompt-injection filtering as one layer rather than a complete agent-security boundary; tool permissions and application-side authorization still matter. - Use API schemas/contracts to constrain what agents may call, and separately authenticate/authorize the identity on whose behalf an agent is acting. ## What to watch - Detailed pricing for AI Runtime Control and AI Firewall relative to dedicated AI gateways/security products. - Independent false-positive/false-negative testing of Fastly’s prompt-injection protections. - Which self-hosted and commercial model providers are supported and how provider-specific features survive routing. - More granular identity/authorization controls for agents acting on behalf of users. - Production case studies measuring latency added by AI inspection and routing at the edge. ## Uncertainties - Fastly’s launch material does not independently establish the effectiveness of AI Firewall against novel prompt-injection attacks. - The release announcement does not provide enough detail to compare total cost with specialist AI gateways and security vendors. - Provider abstraction can hide behavioural differences between models even when API routing succeeds. - Fastly’s machine-generated and AI traffic growth figures describe traffic observed on Fastly’s own network, not the whole internet. ## Sources - [Fastly Launches AI Firewall and AI Runtime Control to Secure and Scale AI](https://investors.fastly.com/news-releases/news-release-details/fastly-launches-ai-firewall-and-ai-runtime-control-secure-and) — Fastly · primary · 2026-09-21T00:00:00+12:00. Official availability, architecture, routing, budget, firewall and API-security details. - [Fastly launches AI firewall & runtime control tools](https://securitybrief.com.au/story/fastly-launches-ai-firewall-runtime-control-tools) — SecurityBrief Australia · specialist_independent · 2026-09-22T00:00:00+12:00. Independent specialist confirmation and security-focused launch context.