Opsin Raises $21M for Enterprise AI Agent Security

James Pham and Oz Wasserman started working together at Abnormal Security before building Opsin with Jeremy Mailen. Their next company initially tackled the sensitive information enterprise AI could expose. As employees began giving agents more tools and authority, the founders followed the problem into the decisions those agents could make.

Opsin announced $21M in funding on October 15, 2026, led by SYN Ventures and Race Capital. The enterprise AI security company plans to accelerate product development, support more regulated businesses and hire. Its expansion puts a practical question in front of security buyers: how should an organization judge an agent whose access is technically permitted but whose actions may exceed its assignment?

The financing backs a product that connects agent identity, purpose, permissions, data and activity. That combination matters as enterprise AI moves deeper into operational work, where approving a connection can also give software the ability to change something another team depends on.

What Opsin's $21M funding announcement establishes

The company-issued release identifies the amount, announcement date and lead investors. It does not name a formal round stage or disclose a valuation. SYN Ventures joins Race Capital in backing a company whose security offering now reaches across several enterprise AI environments.

Opsin previously announced a $7M Seed in April 2025, led by Race Capital with CapitalX, Lockstep Ventures and angel investors participating. The latest announcement does not provide a reconciled lifetime funding total. The historical Seed establishes an earlier financing milestone without supplying the accounting needed to turn the two announcements into a cumulative figure.

The founders' backgrounds help explain the route into this market. Pham, co-founder and CEO, worked on machine learning at Abnormal Security, while co-founder and CPO Wasserman focused on product and security. Mailen, co-founder and CTO, brought engineering experience from Trifacta and Divebell. The current leadership page lists all three in those roles.

For an enterprise buyer, that combination puts product judgment alongside threat detection and data engineering. The commercial assignment is to make an agent's place in a business understandable enough that the security team can approve useful work and explain an intervention when its scope changes.

Why an agent's assigned job changes the security question

Opsin's Agent Intent release, published June 30, describes classification at the time an agent is provisioned. The feature examines the audience, topic, action, goal and guardrail associated with its declared purpose. Opsin says the classifications use model inference with reasoning and confidence signals, giving security staff a view of how configuration fits the proposed job before execution begins.

That is an assessment of available evidence about purpose. It gives a reviewer a starting point for a decision that a list of permissions alone leaves unfinished: whether this particular agent needs this particular reach. An agent serving a restricted internal team and one answering a wider audience can create different exposure even when they connect to the same repository.

Opsin's product description also explains how an agent operating with its owner's credentials can expose information to people who lack direct access themselves. Its Risk Graph combines agent context with changing tools, access and activity. The company describes responses including quarantine, blocking an action and notifying owners about remediation.

The operating consequence is shared responsibility. Security needs enough context to identify a problem, while the person who built the agent may know why the workflow exists and what interrupting it would disrupt. Software that brings those perspectives into the same decision has a more demanding job than producing another inventory of connections.

Cross-platform coverage makes comparison part of the product

Opsin announced Claude Enterprise support in March 2026, alongside coverage for Microsoft Copilot, Copilot Studio, ChatGPT Enterprise and Google Gemini. That history shows the multi-platform strategy predates this funding announcement. The current financing gives the company resources to develop a product already working across different enterprise AI systems.

In his funding commentary, Pham describes building a common agent model because the platforms classify agents differently. Opsin uses that model to connect their purpose and configuration with runtime behavior. The company says it connects through APIs and can deliver an initial risk assessment within 24 hours.

For the buyer, comparison is part of the value. A security team should be able to discuss the reach of agents across departments without beginning each conversation with a different platform's vocabulary. A common model can help that discussion, provided it preserves the details that determine which actions are appropriate in each environment.

Build-time classification and runtime monitoring address different moments in that process. Configuration supplies a proposed boundary; activity supplies evidence of how the agent operates. Both can change as employees add tools, connect data and widen the audience, giving the security team an ongoing relationship with an automation that may have begun as a small departmental convenience.

Where Opsin plans to take the new capital

Opsin says it serves nearly 20 enterprise customers, naming Culligan, Barry-Wehmiller, Wellstar Health System and UiPath in its announcement. Those are company-reported customer relationships and a company-reported count. They place its sales effort in organizations where enterprise data, established workflows and the people responsible for approving AI already have to coexist.

The founder's funding statement identifies product development, broader regulated-enterprise support, an expanded protection layer and recruiting as priorities. The live careers page lists engineering, sales and customer-success positions. That mix suggests implementation and customer adoption are part of the work alongside development, without establishing a hiring target or budget allocation.

Opsin is asking security teams to carry a clearer explanation into the next AI approval meeting. An inventory can identify an agent and a permission review can describe its access; the business still needs to decide whether the proposed work justifies that access as the workflow develops. Pham, Wasserman and Mailen now have additional capital to bring that conversation into more of the systems their customers are putting agents to work inside.