Product1 publisher2 min readPublished
Arcjet puts a policy check in front of every action an AI agent takes
Arcjet's new runtime security takes agent activity in through the OpenTelemetry pipelines platform teams already run, then evaluates each tool call against Open Policy Agent rules before it executes and again after.
The Product Desk · Product desk
What happened
- Arcjet Labs launched agent runtime security, an offering that tracks what AI agents do once they are running inside production systems.
- The product gives security teams a list of which agents are active and logs every run for later review, with each action available to be checked against policy first.
- Activity reaches Arcjet through the OpenTelemetry tooling a platform team already runs, with no code changes or host agent, and Claude customers can route through Anthropic's Compliance API instead.
- Teams can write rules to detect prompt injection or redact personal information before it leaks, and bot detection and rate limits are available as well.
- Arcjet says more than 500 production applications now run its software, three years after David Mytton founded the company.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Someone has to decide what a denied action looks like to the customer sitting on the other end of the workflow, and that is a product call the security team cannot make by itself.
- constraint A security team can start collecting agent traffic without asking anyone for engineering time, but nothing gets blocked until the application team wires up the returned decision, so enforcement moves at their pace.
- capability Reviewing what an agent did after the fact stops being a log-correlation project for whoever is on call.
- exposure The audit trail holds inputs and policy evaluations from every execution, so a company that adopts this takes on a new store of sensitive records with retention questions of its own.
An agent working a support ticket decides a customer is owed money and processes the refund [3]. Under Arcjet's launch, the ceiling on that refund sits in a policy file. Policies run on Open Policy Agent and can be changed without redeploying the application, so a team can cap the dollar value of any refund an agent issues without waiting on a release [8].
The checks run on both sides of the action. Each policy check happens before the action executes, Arcjet returns a decision, and the application can then halt the call, ask a person to approve it or send the agent an explanation [10]. After the call, a second check verifies the result before the workflow moves on [11]. Two checks per action means a ten-call agent run puts twenty decisions in the path of one workflow [17].
Founder and chief executive David Mytton said a risky outcome can build up over "a series of steps that look perfectly reasonable on their own" [4]. The company says its policy controls can act at what Mytton called "machine speed" [5]. The workflows he is describing are mundane ones that start from an email or a code commit and then move across several systems with nobody checking each step [3].
The buyer here is a platform or security team that already runs OpenTelemetry and has agents touching money or customer records. For the regulated version of that team, some controls, personal data detection included, run entirely inside the customer's own process, so that data never leaves the environment [13]. SiliconANGLE's report does not state a price.
The larger framing, that watching agents in production is now its own product category, is more than a single vendor launch can carry. The report covers one company and its funding: Plural led an $8.3 million Series A last October, and Andreessen Horowitz, Seedcamp and former Twilio chief technology officer Ott Kaukver are also backers [16]. The integration list says more about who is expected to buy, naming the Claude Agent SDK, the OpenAI Agents SDK, LangChain, Mastra and Microsoft's Agent Framework among others [12].
The test for a buyer is an incident one of its agents already caused, or nearly caused, and two questions about it. Whether a decision returned before execution would have stopped it, and whether the traces the team already emits carry the tool-call parameters that decision would have to read. A team whose OpenTelemetry instrumentation was built to chase latency has instrumentation work to do before it has policy work.
What to watch
- Whether Arcjet publishes latency figures for the pre-execution check, given it sits in the path of every tool call.
- Whether teams that instrumented OpenTelemetry for performance find their traces already carry tool-call parameters, or have to add them.
- Whether the Anthropic Compliance API route stays a single-vendor path or other model providers expose comparable hooks.