Arcjet launches in-code controls for AI agents taking real-world actions

Arcjet launched agent runtime security on September 17, a product from founder David Mytton that observes, enforces and audits AI agent actions before they execute, available now with a free 15-day trial. The product ingests OpenTelemetry events without application-code changes and supports Anthropic's Claude Compliance API, while developers place an Arcjet guard before tool calls so centrally managed Rego and Open Policy Agent policies return an allow or deny decision before side effects such as refunds, emails or database writes occur. Mytton, who founded Server Density in 2009 and later co-founded the Console.dev developer-tools newsletter before starting Arcjet in 2023, said "A risky outcome can develop across a series of steps that look perfectly reasonable on their own.

Arcjet launches in-code controls for AI agents taking real-world actions David Mytton's third startup connects agent traces across sessions and can block tool, API and database calls before execution. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published Primary source: PR Newswire https://www.prnewswire.com/news-releases/arcjet-launches-agent-runtime-security-for-production-ai-agents-302882668.html Why it matters AI agents can move money, expose data and alter production systems. Arcjet is betting that deterministic checks inside application code can control those actions before damage occurs. In a September 17th announcement https://www.prnewswire.com/news-releases/arcjet-launches-agent-runtime-security-for-production-ai-agents-302882668.html?ref=runtimewire , David Mytton @davidmytton https://x.com/davidmytton?ref=runtimewire launched Arcjet's agent runtime security product, extending his code-first security thesis into production AI agents that can send email, issue refunds and modify databases. Arcjet https://arcjet.com/?ref=runtimewire says the product gives engineering and security teams three connected functions: observe agent activity across a workflow, enforce policies around consequential actions, and preserve an audit trail showing what happened. The product is available now with a free 15-day trial, according to Arcjet's launch post https://blog.arcjet.com/introducing-agent-runtime-security-observe-enforce-audit/?ref=runtimewire . Mytton has been working toward this enforcement model across three startups. He founded cloud-monitoring provider Server Density in 2009 and ran it until its 2018 merger with StackPath https://www.globenewswire.com/news-release/2018/05/14/1501621/0/en/StackPath-and-Server-Density-Merge.html?utm source=openai&ref=runtimewire , where he subsequently led product engineering across edge-computing and security products. He later co-founded the Console.dev developer-tools newsletter before starting Arcjet in 2023. The thread through those projects is operational context. Monitoring tools needed to show developers what production systems were doing. Arcjet applies the same premise to security decisions, placing controls inside the application paths where an agent turns a model response into a side effect. "A risky outcome can develop across a series of steps that look perfectly reasonable on their own," Mytton said in the launch announcement. Arcjet's answer is to connect those steps and evaluate them as one workflow. Security at the action boundary The observe layer can ingest OpenTelemetry events without an application-code change, allowing Arcjet to assemble prompts, tool parameters, identities, session metadata and security decisions into a sequence. Arcjet also supports pulling activity from Anthropic's Claude Compliance API. This gives security teams an inventory of agents and applications, along with individual runs that can be inspected step by step. That matters when an agent starts from an email, retrieves information from a database and schedules a payment through another system. Reviewing each event independently can hide the outcome created by the full chain. Enforcement requires a closer integration. Developers place an Arcjet guard immediately before a tool call or another action boundary. Security teams can then write centrally managed policies using Rego and Open Policy Agent, while application code receives an allow or deny decision before the side effect occurs. A refund tool, for example, can pass the authenticated user's identity, role and proposed refund amount into Arcjet. The application can block the refund, send it for human approval or explain the denial to the agent. Policies can also restrict email recipients and attachments, constrain web-fetch tools to approved URLs, detect prompt injection, redact sensitive information, enforce rate limits and cap usage. Arcjet's agent guard documentation https://docs.arcjet.com/guards?ref=runtimewire makes the architectural tradeoff explicit: the application owns the final enforcement point. That supplies business context a network firewall or model gateway may not have, including the actor, target resource and exact tool arguments. It also means engineering teams must cover the action boundaries that matter. An unwrapped tool or workflow step remains outside that enforcement path. Arcjet is trying to reduce that integration work with native support for the Claude Agent SDK, Claude Managed Agents, OpenAI Agents SDK, LangChain, LangFuse, Strands, Mastra and Microsoft's Agent Framework. Its JavaScript, Python and Go SDKs can also protect queue jobs and background workflows that never receive an HTTP request. After an action, Arcjet records the inputs, policy version, decision and execution context. Arcjet says customers can retain that evidence in its cloud, a single-tenant deployment or a private VPC, or send logs into an existing security system. Some checks, including certain sensitive-information detection, can run inside the customer's process. A three-year thesis becomes an agent product Mytton founded Arcjet around a straightforward complaint: security products were commonly purchased and configured separately from the applications they protected. Developers knew the application logic, while security teams owned the threat model. The controls often arrived after the feature had already shipped. In an April 2026 post https://blog.arcjet.com/developers-can-finally-own-security/?ref=runtimewire , Mytton argued that routine protections such as rate limits, bot rules, prompt-injection checks and budget caps should live in the same pull request as the feature. Agent runtime security carries that argument into workflows where the most important operation may happen several calls after the initial prompt. Arcjet announced a $3.6 million seed round https://blog.arcjet.com/launching-the-future-of-developer-security-seed-funding-from-a16z/?ref=runtimewire led by Andreessen Horowitz https://a16z.com/?ref=runtimewire in September 2024, with Seedcamp https://seedcamp.com/?ref=runtimewire and angel investors participating. An $8.3 million Series A https://www.globenewswire.com/news-release/2025/10/08/3163261/0/en/Arcjet-Raises-8-3m-to-Tackle-Surge-in-AI-Powered-Cyberattacks-and-Malicious-Bots-with-In-Code-Defense.html?ref=runtimewire followed in October 2025, led by Plural https://www.pluralplatform.com/features/why-we-invested-in-arcjet?ref=runtimewire and Ott Kaukver. Andreessen Horowitz, Seedcamp, Jeff Lawson and Feross Aboukhadijeh also participated, bringing Arcjet's announced funding to $12 million. A well-funded security race Agent runtime security has become a busy market as software agents gain access to payment systems, internal data and production infrastructure. Noma Security https://noma.security/solutions/agent-behavior-runtime-control?ref=runtimewire markets session-wide behavioral monitoring and policy enforcement across gateways, hooks, SDKs and direct APIs. Rein Security https://reinsec.io/?ref=runtimewire uses a sidecar to observe agent execution, while larger security vendors are adding agent controls to broader enterprise platforms. The financing gap is substantial. Austin-based HiddenLayer raised a $100 million Series B https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security?ref=runtimewire on September 2nd to expand a platform that includes agentic runtime protection. Onyx Security launched in March with $40 million for an enterprise AI control plane. Arcjet's smaller capital base makes architectural focus important. Mytton is betting that developers will accept an in-code dependency because it offers precise control at the moment an action occurs, while security teams retain central policy management and audit evidence. Rivals are betting on broader discovery, behavior analysis and enterprise control planes. Arcjet's public materials establish the mechanics through documentation and workflow examples. Named deployments of the new product and independent measurements of enforcement latency or error rates are not part of the launch evidence. Arcjet's public pricing https://arcjet.com/pricing?ref=runtimewire starts at $25 per month per application, plus usage charges, with higher tiers at $299 and $799 per month per application. The product's practical test will be whether teams can add guards broadly enough to matter without turning every new tool call into a security integration project. If Arcjet clears that bar, Mytton's long-running belief that security belongs next to application logic gives Arcjet a credible position in a category attracting far larger checks.