Two agent-security rounds landed within 48 hours. Read them together, because they disagree about where the control point is.
Opsin: $21M, API-first. Opsin announced $21M on Oct 15, led by SYN Ventures and Race Capital. Stage and valuation: Undisclosed. It connects by API to enterprise AI platforms, including Microsoft 365 Copilot, Copilot Studio, Claude Enterprise, ChatGPT Enterprise and Gemini, and normalizes their agents into one model. Its Risk Graph ties each agent’s purpose, identity, data access and runtime behavior together and flags drift. That means an agent reaching outside its role, or returning sensitive data to a user who shouldn’t see it. The response options are to quarantine the agent, block the action, or tell the owner what to fix. Opsin says a first agent risk assessment arrives within 24 hours (per Opsin).
Pluto: $34M Series A, endpoint-first. Pluto came out of stealth on Oct 13 with $42M in total funding. Pluto says it secures more than 1.5 million endpoints and has reached multi-million-dollar ARR (per Pluto). TechCrunch breaks the funding into an $8.5M seed led by TLV Partners and a $34M Series A led by Jump Capital, with Skinos and Ansa Capital participating. Valuation: Undisclosed. Pluto deploys sensors on devices, often through existing EDR or MDM tooling. It watches the AI tools, skills, extensions and MCP servers employees run, and it can step into a live agent session to block an action or inject context that makes the agent ask for confirmation (per TechCrunch).
Why the split matters. Opsin’s agents run server-side. They live inside Copilot Studio or ChatGPT Enterprise, hold connectors and act with whatever permissions they were granted. You won’t see those from a laptop. Pluto’s agents run on the laptop: a coding agent, a local MCP server, a browser extension, a finance analyst’s vibe-coded app. You won’t see those through a SaaS admin API. Most enterprises have both kinds. Few will pay for two products to cover them for long.
The practitioner test. Before you take either demo, write three lists. First, which AI platforms expose an admin or audit API you can actually license. Second, which agents run locally under a developer’s or analyst’s token. Third, which identities and service accounts those agents use. The first list is API territory. The second needs a sensor. The third is where the incident will happen, because an agent can only overshare what its identity can read. That last point is why Thursday’s other headline matters.
The exit map just got smaller. On Oct 15, Proofpoint signed to buy Varonis for about $6.3B. Both companies had been buying AI-security startups. Proofpoint took Acuvity in February, and Varonis paid $150M for AllTrue.ai (per SecurityWeek). Those were two obvious buyers for an agent-governance startup. They’re about to be one buyer. Microsoft sells Copilot and the controls around it. The big endpoint and network platforms have their own AI-security lines. The acquirer list is real, but it’s getting shorter while the startup list keeps getting longer.
Our skeptical read. Neither company published a valuation. Pluto’s headline metric is endpoints, which measures coverage, not revenue. Opsin’s is speed to first assessment. Both are good wedges. Neither is a moat yet. The moat is the data model. Whoever builds the agent-to-identity-to-data graph that a SOC trusts enough to auto-quarantine on will win the category, or get bought by whoever is losing it.
What we’d ask in the bake-off. Show me an agent you quarantined in production, and tell me what broke. Show me how you resolve one human running five agents under three identities. Show me what you see when the agent runs on a contractor laptop we don’t manage. Show me the audit trail when your model closes a finding on its own. The answers will sort platform from feature fast.
The read. $55M in two rounds says investors think agent security is a standalone budget line. Thursday’s $6.3B deal says the buyers think it’s a feature of data access. Both can be right for about two years. After that, the control point with the better identity graph keeps the budget.