Opsin raises $21M to give security teams visibility into employee-built AI agents
A startup with roughly 20 employees aims to help security teams see what employees' self-built AI agents are actually doing. The company's own figures — which are not independently verified — suggest that most misaligned agent tools have completely unrestricted access, and that almost no agents require human approval.
Opsin, an AI governance and security company, announced on October 15, 2026 that it has raised $21 million in a funding round led by SYN Ventures and Race Capital. The money will go toward expanding the company's sales and engineering team; Opsin has around 20 employees, according to Business Insider. The round comes at a time when enterprises are rapidly letting employees build their own AI agents in tools such as Microsoft Copilot, ChatGPT Enterprise, Claude and Gemini — often without security departments having visibility into what the agents are actually allowed to do.
What Opsin does
Opsin's platform connects an agent's purpose, permissions, data access, tools and activity to build a complete, context-based picture of how and why the agent operates, the company explains in its press release. The idea is that security teams can then detect behavior that falls outside what was intended — and, ultimately, constrain the agents.
The problem Opsin points to arises because AI agent building has been democratized inside enterprises. When a finance manager can build an agent in Copilot or Claude on their own, a gap opens between who creates the agent and who is responsible for security. The security team sees the tool the agent is connected to, but rarely which permissions the agent has actually inherited from its builder.
The company's own numbers — with caveats
In an analysis of AI agent activity across its enterprise customers, Opsin found that 93 percent of misaligned tools connected to agents had full, unrestricted access to the systems they were connected to. Only one in 200 connected agents required human approval before taking actions, the company writes in the press release.
It is important to emphasize what this is: Opsin's own analysis of its own customer base, presented in a funding announcement. Neither the methodology, the sample size, nor the definition of "misaligned" has been published, and the figures have not been independently verified. That a company selling agent monitoring finds serious security gaps in agent use should be read with that interest in mind. At the same time, the figures point in the same direction as a broader pattern: enterprises have had time to give agents broad access, but not to build controls around them.
A concrete example
The company's founder and CEO James Pham describes, for Business Insider, a case from an enterprise Opsin has worked with: an agent created by a CFO accidentally granted other employees top-level access to financial information. The agent had thus inherited the executive's extensive rights — without anyone actively deciding that ordinary employees should see such data.
Pham frames the monitoring problem with a comparison: "Agents are actually more like humans than you think," he told Business Insider. "And the reason is that you don't understand what's going on in a human head, but you can judge them by their actions." The point is that you don't need to understand a model's inner logic to monitor an agent — it's enough to see what it actually does, with which tools and which data.
Founders and customers
Pham, who was educated at MIT, and co-founder Oz Wasserman met at the security company Abnormal Security and both left there in 2022. According to Pham, they spoke with hundreds of people before launch, and concern about sensitive data was the most common theme that came up.
Opsin itself states that it already serves nearly 20 enterprise customers, including Culligan, Barry-Wehmiller, Wellstar Health System and UiPath, across healthcare, finance, manufacturing and other regulated sectors. The same caveat applies here: this is the company's own stated customer list, and the named customers are not themselves quoted or confirmed in the source material as users.
Investor perspective and a market in motion
Glenn Chisholm, partner at SYN Ventures, explains the investment in the press release: "Enterprises are racing to adopt AI agents, but there is still a fundamental trust gap around giving them access to sensitive data."
The round fits a pattern of investors pumping significant sums into startups focused on AI agent security and governance — a field where several players have recently raised capital. The investors' logic is that agent adoption creates a new attack and error surface that traditional security tools are not built to cover: classic systems monitor users and machines, but an agent functions more like a semi-autonomous in-between that acts on behalf of humans.
What remains to be answered
Several central questions are unanswered in the available material. The company's valuation after the round has not been disclosed. Nor has the type of round been stated. Opsin's figures on data access and human approval lack a methodology description and independent confirmation, and it is unclear how representative they are beyond the company's own customer base — which consists of enterprises that have already bought an agent-monitoring product, which in itself may skew the results.
The most open question is still the market: if Opsin's fundamental claim holds — that enterprises genuinely lack visibility into their own agents' permissions — it will put pressure both on the AI platform vendors, which are increasingly building their own control features, and on a growing cohort of competing governance startups. For Opsin, the next few years are about proving that an independent monitoring layer is something enterprises will pay for, and not just something the platforms themselves will solve.