Analysis
Opsin raised $21 million led by SYN Ventures and Race Capital to build centralized visibility and control over the AI agents employees are creating across Microsoft Copilot, Claude, ChatGPT Enterprise and Gemini, according to Pulse 2.0. The platform links each agent's purpose, identity, permissions, data access and tool use back to a central policy layer, so security teams can flag agents operating outside their intended scope before they cause damage rather than after.
Founders James Pham (CEO), Oz Wasserman and Jeremy Mailen built the company after Pham and Wasserman's prior stint at Abnormal Security, an email-security firm that made its name detecting attacks that looked legitimate on the surface. That background shows up directly in Opsin's pitch: the company's own analysis of customer AI agent activity found that 93% of misaligned agent tools had full, unrestricted system access, and only one in 200 required human approval before acting. That's less a product claim than a liability number -- it describes how little oversight most enterprises currently have over tools their own employees built.
“That's less a product claim than a liability number -- it describes how little oversight most enterprises currently have over tools their own employees built.”
Opsin already counts nearly 20 enterprise customers in regulated industries, including Culligan, Barry-Wehmiller, Wellstar Health System and UiPath, which signals this isn't a speculative future-risk category but a purchase healthcare, manufacturing and financial services buyers are making now. That puts Opsin in a crowded and fast-moving AI-agent-security field: Harmonic Security raised $50 million for workforce AI governance and Reach Security raised more than $50 million for AI agent defense, both within the same two weeks Opsin announced its round. None of the three has disclosed revenue, so right now the signal investors are reading is enterprise logo velocity and the size of the problem Opsin's own data describes, not unit economics.
The open question is whether centralized agent-governance platforms become a durable standalone category or get absorbed into the identity and access management tools enterprises already run, such as Okta or Microsoft Entra. Opsin's bet is that AI agents are different enough -- dynamic permissions, unpredictable tool use, non-human judgment calls -- to need a purpose-built layer rather than a plug-in to existing IAM. The 93% unrestricted-access number it's citing from its own customer base is the strongest evidence for that bet, assuming it holds up as more enterprises actually audit what their AI agents can touch.
