Security teams were built to watch people and servers. The newest employees at many companies are neither, and they arrive holding credentials.
Opsin has closed $21M to police those workers. SYN Ventures and Race Capital led the round, which the startup will spend widening a platform that maps what each AI agent is supposed to do, what it can reach, and how far its actual behaviour has drifted from the brief.
The gap it aims at is permission, not malware. In a review of agent activity across its customers, Opsin found 93% of misaligned tools wired into agents had unrestricted access to the systems they touched. Only one in 200 called for human sign-off before acting.
A registry for borrowed authority
Opsin links agent purpose, identity, permissions, data reach and tool use so security staff see who authorised an agent, which model sits behind it and what it can do. Coverage spans Microsoft Copilot, ChatGPT Enterprise, Claude and Gemini, plus the long tail of internal assistants employees spin up quietly.
James Pham and Oz Wasserman, both formerly of Abnormal Security, founded the company with engineer Jeremy Mailen, betting that agent sprawl would become a governance problem before it became a breach headline.
Paying customers, thin approvals
Nearly 20 enterprises already use it in healthcare, financial services and manufacturing, including Culligan, Barry-Wehmiller, Wellstar Health System and UiPath. That mix suggests buyers worry less about frontier model risk than about ordinary staff handing company data to an assistant that never asked.
The line between an observability dashboard and a control plane is thin, and identity incumbents are moving into the same territory. Opsin’s answer is context: an agent doing something unexpected inside its own role looks different from one that should never have had the key.
