Arcjet Labs has launched agent runtime security, a product that watches what AI agents actually do once they are out of demos and running against real systems. Security teams get an inventory of every active agent, and each action can be checked against policy before it executes, with full runs logged for later review. SiliconANGLE covered the launch exclusively.
Highlights#
The problem Arcjet targets grows out of production agent workflows. Agents now write to databases, answer support tickets, and process refunds, with chains of steps that start from an email or a code commit and cross several systems before anyone notices. Arcjet founder and CEO David Mytton argues that a bad outcome can assemble itself from steps that each look reasonable in isolation, so the product stitches prompts, tool-call parameters, and security decisions from separate sessions into one workflow an investigator can step through action by action.
Ingestion requires no code changes: platform teams can pipe agent activity through the OpenTelemetry tooling they already run, and Anthropic's Compliance API is a second route for Claude customers. Policies run on Open Policy Agent and can change without a redeploy, so a team could cap the dollar value of refunds an agent issues without waiting on a release. Arcjet returns a decision before each action executes: halt it, escalate to a person, or send the agent an explanation. A second check after the call verifies the result. Built-in rules cover prompt injection detection, personal data redaction, bot detection, and rate limits, and native integrations include the Claude Agent SDK, the OpenAI Agents SDK, LangChain, Mastra, and Microsoft's Agent Framework. More than 500 production applications now run Arcjet's software, three years after Mytton founded the company, backed by an $8.3 million Series A led by Plural.
Why it matters#
Payments are the sharpest version of this problem: an agent that can move money needs policy checks before the transaction, not incident reports after. Runtime controls like these pair naturally with escrow and on-chain reputation, which apply the same discipline at settlement time.