The only two figures in the announcement that can be divided into each other are the round and the logo count, so: $100m against more than 50 new platform customers signed over the past year is about $2m of fresh capital per customer added [1][4][15], a ratio that speaks less to unit economics than to a round priced on the year ahead rather than the one behind.
The year behind is where the mismatch sits, or rather where the more interesting version of the story sits. The tenfold ARR growth was earned securing generative, predictive and model-layer systems [4], while Agent Harness Security, the module that extends runtime protection to autonomous coding agents that write, review and ship code with much less human oversight than traditional development tools, was announced on the same day as the money [3][13][16]. So the growth being cited underwrites a product with no revenue history at all.
The demand evidence is also the vendor's own: HiddenLayer's 2026 AI Threat Landscape Report has 96% of organisations calling AI critical to core operations and nearly a third unable to say with certainty whether they have experienced an AI-related breach [10]. That is a genuine measurement gap and a sales deck at the same time, and I would treat it as the second until somebody who pays claims counts it.
The cap table reads like distribution rather than pure return. Delta-v led, with Ten Eleven Ventures, Morgan Stanley, Microsoft's M12 and Booz Allen Ventures alongside [1], and HiddenLayer says its customers include some of the largest names in securities brokerage, banking, insurance, government and the US defense and intelligence communities [5] (neither announcement says whether any investor is also a customer [18]). Ventureburn's account adds that proceeds go to sales and distribution, a new chief revenue officer in Mike Gesnaldo, channel partnerships and EMEA expansion [11], the spending profile of a company with a built product still chasing quota.
This plays out one of three ways. Agentic runtime security becomes a standing line with its own renewal cycle, in which case a 104-patent position, 39 granted and 65 pending, plus the Adversarial Prompt Engineering taxonomy and standards work at CISA, MITRE, NIST, OWASP and OpenSSF is a real moat [7][8][14]. Or the harness and model vendors ship equivalent controls natively, and the module compresses into a checkbox bundled with the agent. Or it stays a defense-and-financials niche where the buyer is a compliance function and growth runs at procurement speed.
This is probably wrong, but I read the middle case as the likeliest, because controls that live inside the agent loop are cheaper for the agent vendor to ship than for a third party to intercept. The counter-thesis is Delta-v partner Dan Williams' argument that traditional tools were built for code and infrastructure, not for models that can be poisoned, hijacked or manipulated through their own inputs [9], and the strongest evidence for it is that one frontier model provider covering more than 700 million weekly users bought rather than built [6] (ventureburn hedges that user figure as reported, since both accounts trace to the same release [19]). What would settle it is renewals on those 50 accounts and whether the coding-agent module ever gets a named reference of its own.
