Outerlimit raises $16m to block rogue AI agent actions

A startup called Outerlimit came out of stealth on 23 September 2026 with $16 million in pre-seed funding for software intended to stop autonomous AI agents from taking harmful actions, according to a report by SecurityWeek.

The company's product is described as a decentralised authorisation layer — software that sits between an agent and the systems it acts on, and decides whether a given action is allowed. SecurityWeek reports that the layer is designed to discover, observe and block harmful autonomous AI actions. Discovery, in this context, means finding the agents already operating inside an organisation; observation means recording what they do; blocking means refusing an action before it takes effect.

The size of the round is unusual for a pre-seed raise, which is normally the earliest institutional funding a company takes, often before a product is generally available. SecurityWeek gives the figure as $16 million. The report does not describe customer numbers, pricing, or a general availability date, and this newspaper has not independently verified the company's claims about its technology.

The pitch reflects a problem security teams have raised repeatedly as agents move from demonstrations into production work. An AI agent that can call tools, browse, send email or write to a database is, from the perspective of the systems it touches, a user with credentials. Unlike a human user, it can be steered by text it reads, and it can act quickly and repeatedly. Controls that assume a person is clicking the buttons do not necessarily apply.

Authorisation layers are one of several approaches vendors are selling against that problem. The appeal is that the check happens at the point of action rather than inside the model, so it does not depend on the agent behaving as instructed. The difficulty, as with any policy enforcement point, is knowing which actions should be refused, and whether the layer can see every path an agent might use. SecurityWeek's report does not set out how Outerlimit makes those decisions in practice, or what "decentralised" means for where the enforcement runs.

Buyers evaluating tools in this category will want to establish the basics before committing: which agent frameworks and model providers are supported, whether enforcement can be bypassed by an agent using a direct API call, how policies are written and tested, and what the logs capture for later investigation. On the evidence available so far, those questions remain open. Outerlimit's funding announcement is a statement of intent rather than a track record.