The Pulse
Fastly Adds Runtime Controls for Enterprise AI Agents
Fastly launches AI Runtime Control, AI Firewall and API Security capabilities for enterprise AI systems. The products provide centralized controls for model access, spending, prompt-injection defense and agent requests to enterprise APIs.
AI.info Team ·
Fastly is adding a control layer between enterprise applications and the AI models they use, betting that companies will need to restrict autonomous systems at the moment they make requests rather than audit those actions later.
The San Francisco-based edge cloud company announced AI Runtime Control, AI Firewall and new API Security capabilities on September 21, 2026. The products target three separate points in an enterprise AI system: applications calling models, users interacting with AI applications, and agents calling internal business APIs.
Fastly presents the launch as an answer to a disagreement now facing enterprise technology teams. Developers want agents to call models and business systems with as little friction as possible; security and finance teams want limits on permissions, token consumption and potentially harmful requests. The company’s products put those controls in the request path.
Fastly’s announcement says machine-generated traffic represented more than 50% of traffic across its network in July and August 2026. It also says AI traffic grew 6.5 times faster than human traffic from January through May, while citing McKinsey’s finding that 93% of organizations are exceeding their AI budgets.
Fastly Puts AI Requests Behind One Control Plane
AI Runtime Control, or ARC, sits between applications and public or self-hosted model providers. Applications send model requests to a Fastly-managed endpoint instead of calling each provider directly. ARC then identifies the virtual key and model, applies limits, substitutes the provider credentials and returns the response.
The design gives administrators a single place to manage provider access without requiring every application to carry raw vendor keys. Fastly says virtual keys can be assigned to applications, teams, developers or sessions, then rotated or revoked without client-side code changes. The company also says customers can continue using their own provider relationships and negotiated rates through a bring-your-own-key model.
ARC records token usage, request volumes and calculated costs against virtual keys and client sessions. It can apply monthly dollar budgets, requests-per-minute limits and tokens-per-minute limits. Fastly says those controls are intended to contain runaway autonomous-agent loops, unexpected traffic spikes and other uses that could exhaust provider quotas.
“To effectively adopt coding agents and implement AI features without losing velocity or eroding existing business resilience, enterprises need control in production, at runtime, without delay, friction, or disruption,” said Kelly Shortridge, chief product officer at Fastly, in the company’s announcement.
AI Firewall Targets Prompt Injection in the Request Path
AI Firewall operates as a security module within ARC. Fastly says it evaluates prompts in the request path to identify and block attacks aimed at AI applications, including prompt injection and other attacks designed for large language model systems.
The product uses Fastly’s SmartParse detection engine and custom guardrails. Fastly’s product documentation describes AI Firewall as an optional layer that can be enabled within ARC, with detection intended to add near-zero latency impact.
Fastly’s product page says the system inspects requests for model-specific threats, applies guardrails to responses and validates outputs for refusal behavior and possible information leakage. The company does not provide independent test results or a measured attack-detection rate in the launch materials.
API Security Limits What Agents Can Do
The third part of the release focuses on the boundary between AI agents and enterprise software. Fastly’s API Security capabilities enforce API contracts across agentic, agent-assisted and conventional traffic. Organizations can monitor or block requests that do not conform to service-specific rules.
That approach addresses a practical problem with agents: their permissions often depend on the APIs they can reach, not only on the model producing their instructions. Fastly says its controls can help prevent agents from requesting unsupported operations, sending malformed payloads or acting outside established API boundaries.
The company describes the feature as a policy layer for agents accessing enterprise APIs, not as a system that determines whether an agent’s broader goal is appropriate. The release also does not claim that Fastly independently verifies an agent’s identity or intent before every request.
Failover and Budgets Bring Operations Into the Product
ARC also addresses reliability and cost management. A virtual key can point to an ordered list of model targets, allowing Fastly to retry a failed request and route it to a backup provider when the primary provider returns errors. The application continues using the same endpoint.
Fastly’s ARC page says administrators can set a monthly budget for each virtual key and choose whether service stops or continues after that limit is reached. The product also offers alerts at 80% and 100% of a configured budget, along with dashboards showing token counts, request volumes and spending.
Those features place model routing, provider credentials, security inspection and usage accounting in one service. Fastly’s separate product post says the company is responding to enterprises that currently use direct provider connections, shared API keys and separate tools for logging, rate limits and cost tracking.
Fastly Extends Its Existing Edge Security Business
The launch expands Fastly’s existing security portfolio rather than introducing an independent agent platform. The company says its AI offering combines ARC and AI Firewall with Bot Management, DDoS Protection, API Discovery, ContentGuard and its next-generation web application firewall.
Fastly’s pitch is therefore about where controls run. Model access, prompt inspection, usage limits and API enforcement can sit on infrastructure the company already uses to deliver and protect application traffic. The approach may appeal to customers that want one network layer for human requests, automated traffic and AI calls, but the launch materials leave pricing, general availability terms and customer deployments unspecified.
For now, Fastly’s concrete offering is a proxy and policy layer: ARC governs the route to models, AI Firewall checks model-directed attacks, and API Security constrains the requests agents send to enterprise services.