Rein Security's $25 million Series A funds a sidecar that blocks AI agent actions as they run

Rein's case for runtime control rests on the gap between what an agent says and what it does. A prompt or a model response may seem benign, and the danger can still come from what the agent does next: opening a record, calling an API, or handing instructions to another system ^[6]. These agents reach business systems and act without a person directing each step ^[3]. Rein argues that judging those actions, and blocking them, takes execution context ^[6].

Matan Bar-Efrat and Netanel Rubin built Rein to monitor what software does in production ^[4]. Their first idea was cyber-insurance software built on security data. Bar-Efrat wrote that insurers showed interest but prospective buyers were slow to pay ^[11]. The founders turned to a problem that cost teams more right away. They had no reliable way to know which vulnerabilities in their live applications were reachable, or which ones actually posed a danger ^[12]. Inspecting an agent's actions and stepping in when one crosses a boundary is the same reachability question, applied to agents ^[14]. Rubin led vulnerability research at Check Point, according to CTech's January coverage ^[13].

Placement is where the adoption cost sits. A sidecar runs beside the workload it watches ^[5]. Coverage therefore depends on deploying it wherever an agent executes, and an agent started outside that footprint goes unwatched ^[5]. In exchange, Rein says, agent data stays in the customer's environment instead of passing through a gateway or proxy ^[5][7]. For agents holding write access to production systems, I think that is the right trade for action-level enforcement. Rein also describes the technology as patented ^[7]. Both are product claims, and neither is an independent measure of coverage or prevention accuracy ^[7].

The field evidence is also Rein's own. Its Agent Breakers research group presented a compromise of a shopping agent at a top-five U.S. retailer at Black Hat USA 2026 ^[8]. The company says it detected and blocked a prompt injection hidden in a PDF that an enterprise onboarding agent opened ^[22]. The announcement does not include enough technical detail to assess the defense in either case ^[23].

Dividing the two growth multiples gives revenue per customer up about 1.6 times since launch ^[25], if both start from the same January baseline. Until Rein gives that baseline, neither multiple says anything about scale ^[24]. The same holds for the claim of thousands of agents executing millions of actions ^[9]. For that figure to tell a buyer anything about load, it would need a period and a per-agent rate. Flex, Swimlane, Lemonade and Dun & Bradstreet are named as users. The names establish customer relationships, not the size of the contracts ^[9].

The funding total is the one figure in the release a reader can add up. Rein's January 2026 seed was $8 million, led by Glilot ^[17]. Add the $25 million Series A and the disclosed rounds come to $33 million ^[20]. Rein puts its total at $35 million ^[10]. It did not disclose a valuation ^[18].

The categories in Gartner's spending forecast include application security, usage controls, governance and gateways ^[16]. Rein is going after the slice covering application security and action control, where its runtime approach competes with products from security startups and larger vendors ^[19].