# What Is Arcjet AI Security? How It Works (2026 Guide) By Osei Tetteh · 2026-11-10 · Source: https://www.activepieces.com/blog/what-is-arcjet-ai-security-how-it-works-2026-guide --- [Arcjet](https://arcjet.com) AI security is a check that runs inside your own application's code, made by Arcjet, which describes itself as "the AI agent runtime security platform." It's not a gateway in front of your traffic or a filter inside a model. ![A stack of application code cards with a small shield-shaped check module slotted inside the stack, and no gateway box in…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/5152b419-5e29-48f4-86ca-629344aaa65f/what-is-arcjet-ai-security-how-it-works-2026-gui-6ae99590.webp) Arcjet makes it, and it's new enough that you should expect the documentation to keep moving. [Activepieces](https://www.activepieces.com) has **747 integrations in its open source monorepo** and each one can be called by an agent, so agents are already being wired into workflows, often before anyone has decided where the guard belongs. Placing it in app code is a real tradeoff: you get a check next to the action, and you own the integration. ## How Arcjet AI security guards your code ### What the maker says it protects against The docs don't describe specific "AI security models," but arcjet.com announces them and shows "Model: Totoro 1.302" on a remote policy example. What the [Arcjet docs](https://docs.arcjet.com) do claim is that it can discover agents and enforce policy across actions, prompts and tool calls, and that's the maker's claim. Most documented events are flaws found before harm was reported: the AI agent incident tracker lists 48 vulnerability disclosures. It also lists **29 incidents, which means real harm happened** often enough to fill a category of its own. ![What kinds of AI agent security events get documented](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/c77d3bbc-347d-4a00-a03a-86b0cea1facc/what-is-arcjet-ai-security-how-it-works-2026-gui-9c28575f.svg "Source: AI agent incident tracker") 11 threat reports is the smallest group in the AI agent incident tracker, so documented attacker activity trails both flaws and damage. Take from it that every claim here rests on two maker-owned sites. The maker is Arcjet, the check runs in the app's code, and the pages read on 2026-11-10 were arcjet.com and docs.arcjet.com. Because both sources are the maker's own, the chart that follows, covering documented AI agent security event types, is the independent check on whether the problem is real. ## What Arcjet AI security does at runtime Arcjet runs as an SDK call inside your own application code. It evaluates each request before the model is called and returns an allow or deny decision your code then enforces. ![A request card moving into an application box where an SDK call block sits before the model box; one arrow with a tick goes…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/d4639275-309c-4d71-82d9-1956ca0f6e54/what-is-arcjet-ai-security-how-it-works-2026-gui-e0fa33e8.webp) ### Where the check runs in your code Right now teams paste user text straight into model calls. A runtime guard changes that by putting a decision point in your handler. The diagram traces a request from user or agent input into your application, through the Arcjet check inside it, and on to the model and back as a response. The allow/deny branch leaves the check itself. ![Activepieces pricing page with four subscription tiers showing costs, features, and call-to-action buttons](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/841ec84e-02e4-4761-aca2-e92f6d457f41/self-host-mistral-ai-enterprise-deployment-guide-c7d7dca9.webp) ![Request path diagram](request-path-placeholder) What follows from that placement is a tradeoff. The check sees your request context and your route logic, which a network gateway can't. **In exchange, you own the integration, and every code path that calls a model must call the check too.** Aimed at AI traffic, the protections the maker's pages describe include: - prompt injection detection on incoming input - sensitive information detection, so personal data is caught before it reaches a model - rate limiting, to cap how much any one caller can spend - bot detection, to separate automated callers from people Each one is a decision your code receives, so least privilege for an agent becomes something you write as rules, not something you hope a filter infers. I'm not asserting anything beyond what those pages state. Detection accuracy and whether model responses are inspected on the way back are questions the pages I relied on don't settle, and the added latency (around 100ms for prompt injection detection, under 1ms for local decisions) is only the maker's claim. Ask the maker, or test in staging, before a rollout depends on any of them. ## How to get access to Arcjet AI security You need a signup, a key and an SDK inside your own codebase. Nothing sits in front of your traffic, so nothing needs approval from a network team. The steps follow [Arcjet's docs](https://docs.arcjet.com), and the maker's own pages at [arcjet.com](https://arcjet.com) describe the product as a runtime platform to observe, enforce and audit agent actions. 1. Create an account through the [Arcjet docs](https://docs.arcjet.com). Use a team address rather than a personal one, so the key outlives whoever set it up. 2. Create a site in the dashboard and copy its key, following the [docs](https://docs.arcjet.com). Treat it like any server-side secret: environment variable, secrets manager, never the client bundle. A separate site per environment keeps staging experiments out of production results. ![A numbered setup sheet with an account signup form, a dashboard page with a site entry, and a key copied onto a small key…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/ac16bceb-7504-4afc-b9b7-185750ec63d6/what-is-arcjet-ai-security-how-it-works-2026-gui-28a2119f.webp) ### How to install Arcjet AI security SDK 3. Install the SDK for your language, using the install page in the [docs](https://docs.arcjet.com). Check the docs for the current list of supported languages and frameworks before you commit, because that list, not this article, is the authority. 4. Add the AI protection rule to your Arcjet configuration, as described in the [docs](https://docs.arcjet.com). The rule lives next to the code that calls your model, so you decide which prompts and which routes it covers. That's the design choice that separates it from a gateway. 5. Test it by sending a known injection attempt and an ordinary request through the same path, as the [docs](https://docs.arcjet.com) walk through. Run it in staging first and note what gets blocked, what slips past and what the call adds to response time. Once those five steps work, you have evidence from your own traffic to set against the maker's claims. ## What Arcjet AI security costs Read the cost from Arcjet's own pricing page, checked on 2026-11-10. Treat anything in this article that sounds like a price as stale until you've re-read it, because this section deliberately carries no figures. That fits the evidence you just gathered: a runtime guard in your own code is priced by how much traffic it inspects, so your measured volume is the input that matters. ### Arcjet AI security published pricing plans The screenshot below shows that shape on a different vendor's page, as an illustration of what to look for: each tier lists its limits and support level. Compare that structure against Arcjet's page, then line up your own request volume against each tier's stated limits. The tier you land on tells you whether the free plan covers a pilot or only a demo. Once you know which tier fits, the question becomes what the page leaves out. ### What Arcjet AI security pricing leaves out Anything missing from the pricing page is unpublished, and I haven't filled those gaps. If the page doesn't state a limit, an overage rate, a retention period or an SLA, don't assume one. Ask Arcjet directly and get the answer in writing before you commit. Recheck the page before any budget sign-off. Plans and limits on a live pricing page change without notice, and a figure copied into a business case today may not match the page next quarter. ## Where Arcjet AI security excels and falls short Arcjet's design gives it real advantages over gateways and model-side filters. Because it's **a new product with no outside testing**, treat those advantages as claims until you have run it yourself. This assessment rests on the maker's pages and design, and we haven't run it. ### Arcjet AI security strengths from its in-code design The usual alternative is scattered if-statements nobody reviews, wired into agents by hand. Arcjet replaces that with a check inside your own code, so it sees the same request, user and tool call your application sees. A gateway sitting upstream sees only traffic, not intent. Placement also dictates reach. Arcjet offers native SDKs and framework integrations, with SDKs for JavaScript, Python and Go. That's broad coverage of the agent tooling people actually use, so adoption is a library import, not a new network hop. But three languages means a team on any other stack finds no documented path in. Activepieces has no native Arcjet integration, but it covers part of that gap. That same action is reachable from Claude, ChatGPT, Cursor or an agent you built, with no second integration. A check service you host and call from an HTTP request step sits in the flow the same way every other connector does. ![Activepieces workflow builder with HTTP request step selected and its settings panel open showing method, URL, headers, and…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9e36e973-de4a-435f-8912-12e666216b2d/self-host-mistral-ai-enterprise-deployment-guide-3f7fe28d.webp) ### Arcjet AI security gaps and unknowns It's young, and the docs describe what it should do, not what independent testers have measured. The wider picture is still thin. The [Blue Radius](https://blueradius.io/research/ai-agent-incident-tracker) tracker's documented agent cases moved from 41 to 34, while those confirmed as real-world incidents rose from 7 to 11. ![Documented AI agent cases: all versus real-world](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/8eda72c9-eef7-4fa8-8774-142e48fa40db/what-is-arcjet-ai-security-how-it-works-2026-gui-108936be.svg "Source: Blue Radius") Fewer cases overall but more of them real suggests the field is **shifting from hypothetical attacks to live ones**. The sample is also small, so no one can yet say how often a runtime guard would have stopped one. That leaves the limits the docs themselves state. It also leaves everything we can't know without a trial: false positives on your traffic, added latency, and how it behaves against an attack nobody has catalogued. A pilot on one low-risk agent would answer those. ## Arcjet compared with established AI security options Where Arcjet differs from the other approaches is placement: **the check sits inside your application's own code**, not in front of it, beside it, or inside a model vendor's service. The table shows how that placement changes the trade-offs, and the main thing to take from it is that **each option fails in a different place**. | | Where it runs | What it inspects | Setup effort | Cost model | Failure behaviour when down | |---|---|---|---|---|---| | Arcjet in-code guard | Your application, calling the maker's service | Prompts and requests your code hands it | SDK added per route or agent | Vendor plan | Your code decides whether to allow or block | | AI gateway | Network proxy between app and model | Traffic passing through it | Reroute model calls | Platform or usage fee | Depends on whether traffic can bypass it | | Provider guardrails | The model provider's platform | Inputs and outputs on that provider | Provider settings | Within provider billing | Tied to the provider's availability | | Open-source library | Your own process or servers | Whatever you wire in | You integrate and host | Free licence, your hosting | Fails with your application | _Prices and plan limits checked against [docs.arcjet.com](https://docs.arcjet.com) and [arcjet.com](https://arcjet.com) on November 10, 2026._ Pricing and behaviour vary by vendor, so confirm each against the maker's current pages before committing. If you want one choke point for every model call, a gateway suits central teams. The cost is that it sees traffic, not your application's context, and it becomes shared infrastructure to keep running. Cheapest to switch on are provider-side guardrails, but they cover only that provider's models. If an agent uses several providers, you end up with several policies that don't match. If you want control and no licence fee, open-source guardrail libraries give you both. You pay in integration, hosting and upkeep, which is a fair trade if you already run that kind of service. ## Connect Arcjet to an Activepieces workflow today With no Arcjet integration in Activepieces, the connection is built from generic steps, and we haven't tested any of these three options. The options differ in effort, running from least to most. This is the simplest route, but only if Arcjet's documentation lists a public endpoint you can call directly, and we can't point to one. Arcjet is built as an SDK inside your own code. So the realistic version is a small service you host that runs the Arcjet check and returns allow or deny. The workflow calls it before any model step and branches on the answer. Because the check sits in your service, it's a gate the workflow can't skip. ### AI step with a custom provider If the AI step can be pointed at a custom or OpenAI-compatible provider, which we haven't confirmed, you would put a proxy at that address that applies the Arcjet check and then forwards the request to the model. The Activepieces core is MIT-licensed and self-hostable, so that proxy can run on the same infrastructure. The workflow stays tidy, but you have now built the gateway, with its hosting and upkeep, that Arcjet's own design avoids. Choose this only if many workflows should be guarded without each one adding a step. ### MCP server Here you would expose an Arcjet check as a tool on an MCP server and let the workflow's agent call it. This is the weakest of the three as a control. The model decides whether to call the tool, so a skipped or misjudged call means no check ran. It suits advisory screening, not enforcement. If a block must hold, use the first option and make the check a fixed step in the flow. ## Frequently asked questions Is Arcjet AI security open source? Partly: the client SDKs that you install in your application are published openly, while the hosted service that analyzes requests is the maker's own product. Because the guard sits in your code, you can read what runs in your process, but the analysis itself isn't something you can self-host from those pages. ### Does Arcjet AI security replace a firewall? No. It decides inside your application, per request, with the context of your own code, such as who the user is and which route or model call is being made. A network firewall, an AI gateway, or a model-side filter works at a different layer. The tradeoff is coverage versus context, so treat Arcjet as an added layer rather than a swap, unless you have already decided your gateway's checks are redundant. Which languages does Arcjet AI security support? It targets JavaScript and TypeScript applications, with Python also documented as an SDK. If your agent code is in another language, you can't drop the guard in directly, so check the current docs for your framework before you plan the rollout. Can I try Arcjet AI security for free? Yes, the maker's pricing page lists a free plan, which is enough to wire the guard into one non-critical route and see what it blocks. Limits and which AI features are included can change, so confirm them on the pricing page before you promise the pilot to anyone. ## Related reading - [Arcjet AI Security Models: Pricing and Features (2026)](https://www.activepieces.com/blog/arcjet-ai-security-models-pricing-and-features-2026) - [What Is NVIDIA OpenShell? A Runtime for AI Agents](https://www.activepieces.com/blog/what-is-nvidia-openshell-a-runtime-for-ai-agents) - [No Code Web Builder: Best Options in 2026](https://www.activepieces.com/blog/no-code-web-builder-best-options-in-2026) ## References - [Arcjet](https://arcjet.com/pricing)