AI Agent Risk Governance: Opsin Raises $21 Million

Opsin raised $21 million for AI agent risk governance software in a round led by SYN Ventures and Race Capital, the company said on October 15.

The money lands as compliance and security teams try to account for software agents that employees build inside Microsoft 365 Copilot, Claude Enterprise, ChatGPT Enterprise and Gemini. Those agents hold credentials, connect to corporate data and take actions without a person approving each step, which is a different control problem from monitoring a user typing prompts.

Opsin sells to enterprise security teams, and its customers include firms in healthcare, financial services and manufacturing, sectors where regulators ask firms to show who can reach which data. The company has nearly 20 enterprise customers, among them Culligan, Barry-Wehmiller, Wellstar Health System and UiPath.

What Opsin raised and who backed it

SYN Ventures and Race Capital led the $21 million investment. Race Capital also led Opsin’s earlier $7 million seed round, which the company used to stop generative AI tools from oversharing sensitive files.

Opsin said the new capital will fund faster product development, expansion among regulated enterprises, further growth of its risk protection layer as agents gain autonomy, and hiring in sales and engineering.

Alfred Chuang, general partner at Race Capital, said the firm backed the team for its security experience and technical depth. “Opsin connects what an agent is supposed to do with what it actually does,” Chuang said.

Glenn Chisholm, partner at SYN Ventures, framed the investment around trust. “Opsin provides enterprises the confidence to embrace AI agents without opening the door to uncontained risks,” Chisholm said.

How Opsin approaches AI agent risk governance

Opsin started with data oversharing in Microsoft Copilot and has widened its scope to the agents themselves. The platform now tracks each agent’s identity, intent, connectors, tool calls and permissions, rather than only the data moving through a chat session.

Platforms label agents in different ways, so Opsin built a standard model that normalizes agents across Microsoft 365 Copilot, Copilot Studio, Claude Enterprise, ChatGPT Enterprise and Gemini. It then scores two things: agent posture, which covers connectors, tool calls and sharing scope, and runtime behavior.

Risky runtime behavior includes an agent reaching beyond its role scope or delivering sensitive data to users whose roles do not match. That context feeds the Opsin Risk Graph, which learns from agent activity to detect new risks.

When an agent breaches policy, Opsin can isolate it, stop the action or alert the agent’s owner to the fix. The platform connects through an API and produces an initial risk assessment across approved AI platforms within 24 hours.

“Security needs to understand what the agent is trying to do, who is behind it, what data it is using,” said James Pham, co-founder and chief executive of Opsin. Pham and co-founder Oz Wasserman previously worked at Abnormal Security. The third co-founder is Jeremy Mailen.

What the agent data shows for risk teams

Opsin’s analysis of its enterprise customers found that 93% of misaligned tools connected to agents had full, unrestricted access to the systems they touched. Only one in 200 required human approval before acting.

Opsin said in June that its 24-hour assessment finds an average of 853 agents per customer, many unknown to security teams. It said customers cut AI and agent risk by an average of 70% within six weeks. The same announcement said Gartner listed Opsin as a Representative Vendor in its 2026 Market Guide for Guardian Agents.

For a compliance team, the practical question is evidence. An inventory of agents, a record of what each one accessed and a log of blocked actions are the kind of artifacts an internal audit asks for. They also resemble the data-driven audit trails that firms already build for other control areas.

Opsin said the aim is to let security teams see every agent, understand why its behavior is risky, rank the issues and hold the evidence behind those decisions. Pham said in June that the company “started by solving data oversharing in Copilot” and is “now securing an entire generation of autonomous enterprise agents.”

The rules behind agent oversight

No single rule requires a tool like Opsin’s. The pressure comes from several frameworks that ask firms to govern AI systems and control access to data.

The EU AI Act, Regulation (EU) 2024/1689, requires providers and deployers of high-risk AI systems to put human oversight and risk management in place. The NIST AI Risk Management Framework, published in January 2023, is voluntary and organizes AI risk work into four functions: govern, map, measure and manage. ISO/IEC 42001, issued in December 2023, sets out requirements for an AI management system that auditors can certify against.

Existing privacy and security obligations apply to agents too. An agent that surfaces personal data to the wrong employee creates the same exposure as a misconfigured file share. Banks are already testing the technology in monitoring work, as RegTech Post reported in its coverage of banks turning to AI agents for AML, fraud detection and risk monitoring.

Where Opsin sits among agent security vendors

The field has drawn large rounds. Zenity raised a $125 million Series C led by Norwest in August 2026. Noma Security raised a $100 million Series B in July 2025. Both sell security for AI agents to enterprise buyers, and Opsin competes with them for the same budgets.

Opsin’s pitch rests on its roots in Copilot oversharing and its depth in the major enterprise AI platforms. It focuses on agents that employees create inside approved tools, rather than on agents that developers build and deploy in custom code.

A separate group of vendors approaches the same problem as compliance automation. Norm AI builds agents that perform compliance work, which makes it a user of agents rather than a control layer over them. The two categories are likely to meet inside the same large firms, where one team deploys agents and another has to supervise them.

Pham expects agents to interact with one another far more than with people within three years, with much of that exchange never reviewed by a person in real time. Opsin’s position is that organizations should oversee agents as they oversee employees, judging them on their actions and the demands of their role.

What comes next for agent oversight

The effect on compliance functions is already visible in how teams are organized. As agents spread, ownership of their risk is shifting between security, privacy and risk groups, a change RegTech Post has followed in its look at how regtech is reshaping compliance teams.

Opsin said it will use the funding to deepen coverage as vendors add agent features. It has already added support for Claude’s Compliance API and for Claude Cowork in 2026. The next marker to watch is how many of its roughly 20 enterprise customers in regulated sectors expand beyond their first approved AI platform.