HiddenLayer raises $100 million to protect AI agents at runtime

From protecting models to controlling actions

HiddenLayer raised $100 million in a Series B. The company itself confirmed the round. The funding is directed towards protecting coding agents and monitoring the behaviour of autonomous systems in production. This reflects a change in threats: testing a model before launch is insufficient if the agent then gains access to tools, keys, internal data or the ability to modify external systems. [1 · HiddenLayer]

The platform is broadening protection: model testing is being supplemented by defences against prompt injection, prevention of tool misuse and blocking of unauthorised actions. Such a runtime control layer needs visibility into the agent registry, must monitor tool calls and must be able to stop an action before it results in publication, a money transfer, a configuration change or a data leak. [1 · HiddenLayer]

Company growth and the limits of the data

HiddenLayer reports more than tenfold growth in ARR and more than 50 new platform customers. It also mentions a customer that develops frontier models and has more than 700 million weekly users. Exact ARR figures, the customer's name and the calculation methodology are not disclosed. These metrics are public company claims, rather than audited accounts, and must be published with that caveat. [1 · HiddenLayer]

Despite the limited disclosure, the direction of demand appears logical. Large buyers want more than a set of static safeguards: they also want an agent registry, runtime controls, a log of reasons, permissions management and an emergency stop. The more autonomy an agent receives, the closer security becomes to essential execution infrastructure. [1 · HiddenLayer]

What basic controls look like

A practical architecture includes tool allowlists, separate credentials, limits on transaction amounts and the scope of operations, human confirmation of critical actions and an immutable audit log. It is important to check not only the content of a model's response but also the actual action, its authorisation context and the ability to roll it back. These are the features that could become part of enterprise procurement standards for agent systems. [1 · HiddenLayer]