The Third Round in Three Weeks: Cymphony, Agent Identity, and the Crowded Race to Govern Machines

Over the past five months, roughly $435 million has poured into a single security category. Three of those rounds closed within three weeks of one another. The most recent is Cymphony — $25 million, led by Sequoia, at a post-money valuation that has only just cleared $100 million. The company was founded in 2023. Its first selling year produced seven-figure ARR. I found the story on a feed labeled as blockchain and Web3 news. There is not a single wallet address in it. No chain, no token, no validator. What arrived was enterprise security coverage wearing a crypto masthead — evidence that "blockchain journalism" has quietly become a distribution label rather than a beat. Strip the label and something genuinely interesting is underneath. Cymphony builds what it calls a workforce graph: a unified map of human identities, machine identities, data, and activity signals inside an enterprise. Its founders come out of Israeli military intelligence — the Talpiot and 8200 pipeline that also produced the founding team of Wiz, and before that a long line of endpoint, checkpoint, and cloud-security companies. The commercial wedge is identity-first. Rather than bolting guardrails onto model outputs, Cymphony tries to govern who — or what — is allowed to touch which data, and to leave a record when they do. The customer list is the most honest part of the story. KKR, Syngenta, Cass Information Systems. A global private equity firm, an agricultural multinational, and a US-listed payments processor. Their common denominator is not enthusiasm for AI. It is exposure: each operates under heavy regulatory scrutiny, each handles dense sensitive data, and each has already lost something to shadow AI — a file pasted into a consumer chatbot, an unapproved tool living quietly inside a corporate SSO session. The investor structure rewards a second look. Sequoia led. A fund described as SMBC Fin Atlas Beyond Fund participated — SMBC being Sumitomo Mitsui, and Japan's regulator having spent two years pushing financial institutions toward demonstrable AI governance. That pairing is not accidental. It is distribution: regulated finance in two jurisdictions, wired in from day one. Digging deep for the truth in the chain is a phrase I usually reserve for on-chain forensics. Here it applies to a permission graph. In 2017 I wrote a static analysis tool to hunt reentrancy bugs in ERC-20 contracts. EthGuard Lite found twelve critical flaws in my own project's codebase. The lesson that stayed with me was not that bugs exist — it was that finding and fixing are separated by a chasm. Every security product is judged on the wrong side of that chasm. Cymphony's public description sits on the discovery side. It surfaces files exposed to AI tools. It surfaces unauthorized AI tool usage. Nowhere in the material is there a claim of inline blocking — of severing a session mid-prompt, or revoking an agent's scoped token the moment its behavior drifts. That distinction is not semantic. Out-of-band detection tells you, hours later, what already happened. Inline enforcement changes what happens. The engineering difficulty, and the contract value, differ by roughly an order of magnitude. The named methodology — a unified graph spanning identity, data, and activity — is a technique, not a new computational paradigm. Data security posture management already catalogues where sensitive information lives. Identity threat detection and response already watches for anomalous credential use. User and entity behavior analytics already scores deviation from baseline. What Cymphony appears to have done is converge three mature categories under a sharper story and a cleaner entry point. That can be excellent product strategy. It is not a moat. The moat question matters because the category has already entered its consolidation phase. Palo Alto Networks bought Protect AI. Cisco bought Robust Intelligence. Check Point bought Lakera and Lasso Security. SentinelOne bought Prompt Security. F5 bought CalypsoAI. Cyera bought Oasis Security for non-human identity. That is at least six exits before Cymphony's Series A — an unusually loud signal that the category is validated and that the independent ceiling is low. Meanwhile the platform vendors compress from above. Microsoft has put agent identity into Entra and shadow-AI discovery into Purview, essentially free to anyone already paying for E3 or E5. Palo Alto, Zscaler, Netskope, CrowdStrike, Varonis and Cyberhaven have folded AI usage governance into existing suites. A buyer with an Entra seat now asks a fair question: what am I paying Cymphony for that I do not already own? The most conspicuous blind spot is the one nobody in the coverage mentions: MCP. The Model Context Protocol has become the connective tissue of agent deployments, and it has also become the year's richest new attack surface — tool poisoning, prompt injection propagating through tool chains, unmanaged MCP servers multiplying inside enterprises that keep no inventory of them. A governance product that does not address MCP runtime is a governance product for the previous architecture. This is where my DAO work keeps intruding on my security work. At Synapse DAO we trained a model on ten thousand historical governance votes and hit 85 percent accuracy predicting outcomes. It stopped one proposal that would have destroyed roughly $5 million in treasury value. But the model never solved the underlying problem: a DAO can simulate a vote and cannot revoke a delegate's authority mid-execution. Identity without runtime enforcement is a forecast, not a government. Audit complete. The soul remains. The obvious contrarian take is that $100 million-plus for a company with single-digit-millions ARR is insane — 33 to 100 times revenue, against 10 to 25 times for CrowdStrike or Palo Alto in public markets. That take is lazy. Private AI security leaders have cleared 40 to 60 times ARR. Cymphony sits at the top of the band, not outside history. The number that deserves scrutiny is dilution. Twenty-five million on roughly a hundred million post-money means the founders sold about a quarter of the company at Series A, where 15 to 20 percent is normal. That is either urgency or a very strong lead investor's pricing power. Either way it is more informative than the headline valuation. Two more things are missing, and their absence is itself data. There is no net revenue retention, no average contract value, no sales-cycle length. And there is no disclosure of whether the seed investors followed on. If they did, the coverage buried a positive signal; if they did not, it buried a warning. Archaeologists of the abstract learn to read the gaps in the record as carefully as the record itself. The forward question is not whether Cymphony survives. It is whether identity becomes the control plane for an economy in which most actors are not human. Every agent that transacts on your behalf needs an owner, a scope, and a revocation path. Financial regulators in Tokyo and Washington are already asking for exactly that. The company that makes agent identity auditable does not sell software — it sells the constitution for machine labor. Whether Cymphony is that company is unresolved. Whether that constitution gets written is no longer in doubt.