Enterprises should give AI a governed path rather than ban it, Armor founder and CEO Chris Drake told a GovWare 2026 session in Singapore — a case for AI governance, not prohibition. His argument, in a talk titled “Once Again. AI Puts CISOs Between a Rock and a Hard Place,” is that AI should complete useful business work automatically within policy — with security leaders setting the conditions for adoption and holding the authority and evidence to enforce them. “With the cloud, the CISO at least got a say. With AI, nobody asked,” he said. “Banning it doesn’t work, and ignoring it is how the breach report gets written.”
The departure point: AI arrives without the cloud transition’s checkpoint. When workloads moved to cloud, security teams could assess a provider and influence what left; AI arrives inside existing browsers, productivity tools and employees’ machines, and an agent with the right permissions can read files, send data and take actions. The risk can travel over a connection the firewall allows, so “the firewall works as configured” is no longer reassuring.
The blind spot
The most concrete part — endpoint visibility — holds up. Drake pointed to two documented gaps: Anthropic states that host endpoint detection and response (EDR) cannot inspect activity inside its Cowork local VM, and Docker documents that processes running inside its local sandbox microVMs are invisible to the host. An employee clicking “Allow” does not settle whether the business approved the access.
Independent research corroborates the pattern: Cloud Security Alliance analysis found agentic command-and-control “looks like a user sitting at a desk” to host tooling, and that leading EDR platforms produced no alerts in a proof-of-concept. GuardionAI describes “phantom workflows” — ephemeral agents that execute and vanish beneath EDR polling — and Rye notes that microVMs trap system calls to a guest kernel, leaving host telemetry nothing to attach to. The gaps Drake cited sit inside a broader limitation of endpoint security as agents move into sandboxes.
The AI governance model
From that gap, Drake argues for a sequence, not a ban. First, data classification moves from nice-to-have to must-have: organisations must know what their data contains, who owns it and how they may use it, because, as he put it, “AI cannot safely use data you cannot govern.” Second, organisations should own the routing policy — approved external models when data policy permits, private execution when required. Third, every authorised action should leave a receipt: records of the data accessed, the policy decisions, the models, the routes, the actions and the outcomes, so an organisation can reconstruct what happened without a person approving every step.
He also flags fourth-party risk: a vendor’s use of another AI provider to process data adds exposure to document. Protiviti’s February 2026 AI Pulse Survey puts a figure on the problem, if imperfectly: 47% of large organisations reported incomplete visibility into employees’ AI-tool use. Drake’s practical test before granting an AI tool access is three questions — what can it touch, where can it send, and can we control what it does.
The frame matters: Drake sells the remedy. Armor, which has served regulated industries since 2009 and says it protects more than 1,700 organisations, built a governed AI platform for its own operations and has productised it as Sovereign AI, demonstrated at GovWare. That does not weaken the argument — the platform vendors themselves document the EDR gap he cites, and independent research corroborates it — but the governance model arrives with a product attached.
What to watch
Three things will show whether the governance-first approach becomes the norm. First, whether endpoint vendors close the sandbox visibility gap with agent-aware telemetry. Second, whether the share of organisations lacking visibility into AI use falls as governance platforms spread. Third, whether buyers treat governance as a checkbox or a discipline: the receipts matter only if someone reads them.
What the talk does and doesn’t establish
It establishes a coherent governance argument, a corroborated endpoint blind spot, a three-step model (classify, route, audit) and questions for CISOs. It does not establish that Sovereign AI delivers on the model, any customer outcomes, or that a single platform can solve a governance problem that spans employees, SaaS tools and shadow agents.
The honest summary: a vendor talk with an unusually verifiable core. Anthropic and Docker document the claim that matters — that host-based security cannot see inside the local VMs where AI agents increasingly run — and independent research echoes it, which makes it a better starting point than most AI-governance rhetoric. Drake’s conclusion — an AI governance answer, not a ban — is one other vendors and researchers are converging on. The open question is the one every governance pitch faces: whether the controls work, and whether organisations read the receipts.
Editor’s Note: This article draws on the Armor release and the company’s published materials, and on independent research by the Cloud Security Alliance, GuardionAI and Rye, plus Docker and Anthropic documentation and Protiviti’s February 2026 AI Pulse Survey. The talk and the governance model are as described by Armor; the endpoint-visibility gaps are corroborated by vendor documentation and independent research, while Sovereign AI‘s efficacy is not independently verified. Armor sells a governance platform. No interview was conducted.