For two decades, corporate security has meant watching the laptop. Every process. Every file write. Every outbound connection. Some AI agents now run in a place that watching can’t reach.

That was the most concrete point Armor founder and CEO Chris Drake made at GovWare 2026 in Singapore, according to a press release Armor Defense Inc. issued on October 14. His wider argument, that security chiefs should approve AI under enforceable rules instead of banning it, also happens to be the sales case for Sovereign AI, the governed AI platform Armor sells. The technical point stands on its own. The commercial angle is worth keeping in view anyway.

The Blind Spot

Most large companies run endpoint detection and response software, usually shortened to EDR. It sits on each laptop and server and flags behavior that looks like an attack.

Drake’s argument: AI agents increasingly run inside a small virtual machine on the employee’s own computer. A virtual machine is a self-contained computer simulated in software, with its own operating system. EDR on the host often can’t see into it.

He cited two vendors by name. We checked both.

Anthropic’s support documentation for Cowork, its desktop agent product, answers the question directly. EDR tools can’t inspect activity inside the Cowork VM, which is isolated from host security tools by design. Sessions that run in the cloud sit outside company endpoints entirely. The page advises customers whose compliance depends on endpoint visibility to account for this before rollout.

Docker says much the same about Docker Sandboxes. Each sandbox is a microVM with its own Linux kernel, and Docker’s isolation documentation states that processes inside it are invisible to the host.

So the claim holds. There’s a catch, though, and the release only touches on it.

The wall is the safety feature. It keeps a misbehaving agent away from the rest of the machine. It also keeps the machine’s own guards out. Neither vendor hides the tradeoff, and both offer other ways to watch. Anthropic points admins to OpenTelemetry, an open standard for streaming activity logs, plus its Compliance API. Docker routes sandbox traffic through a proxy on the host that enforces an allow list. Monitoring shifts from the endpoint to platform logs and the network. Drake’s release concedes as much, noting that platform telemetry and network controls remain important.

He added one sharper line. An employee clicking “Allow” on a permissions prompt, he said, doesn’t settle whether the business approved that data access.

“Nobody Asked”

Drake framed AI as a rerun of the cloud migration, with one difference.

“With the cloud, the CISO at least got a say. With AI, nobody asked. Banning it doesn’t work, and ignoring it is how the breach report gets written,” he said.

During the cloud shift, security teams could vet a provider before data left the building. AI arrives inside browsers and office tools employees already use. Depending on its permissions, an agent can read files, send data and take actions over a connection the firewall already allows.

Then there’s the supply chain. If a software vendor sends your data through another company’s AI model, that’s fourth-party risk: your vendor’s vendor. Drake said companies need evidence of which providers touch their data, where it’s processed and what’s retained.

The Survey Behind the Shadow AI Claim

The release cites Protiviti’s AI Pulse Survey, and the figure checks out. Protiviti, a consulting firm owned by Robert Half, surveyed about 345 executives, board members and IT leaders in February 2026 and published the results in May. Forty-seven percent of large organizations said they lack full visibility into the AI tools employees use. Sixty-five percent reported problems with “shadow AI,” meaning tools used without proper oversight. Only four in 10 organizations had a formal AI governance framework.

One caveat. Protiviti sells AI governance consulting. It’s a corporate survey, not independent academic research.

What Drake Wants CISOs to Do

His prescription comes in layers. Classify data first, so systems know what’s sensitive and who owns it. “AI cannot safely use data you cannot govern,” he said. Then let the company, not the individual employee, own the routing rules: approved outside models where policy allows, private execution where it doesn’t. Log every data access, model, route and action, so an incident can be rebuilt afterward without a person signing off on each step.

On governance, he wants CISOs to hold a written mandate to approve, restrict or pause AI use. Each risk exception should have a named executive owner and an expiration date.

And before any AI tool gets access, three questions. What can it touch? Where can it send? Can we control what it does?

The Product Behind the Talk

Armor, founded in 2009 as FireHost, is privately held and has no public stock. The company says more than 1,700 organizations in over 40 countries rely on it. That figure is Armor’s own.

According to the release, Armor worried that sending customer data through public AI services wouldn’t meet its own customer commitments, so it built a governed AI platform for internal use and now runs its business on it. That one internal deployment is the only usage the release describes. The platform became Sovereign AI, which Armor is demonstrating at its GovWare booth through October 15. The release names no outside customers, deployment counts or independent test results.

Why It Matters

Armor didn’t uncover the EDR gap. Anthropic and Docker document it themselves, in plain language. What Drake did was put it in front of a room of security buyers during Singapore International Cyber Week and argue that it changes who should be making AI decisions.

Whether Sovereign AI is the right answer is a separate question, and the release can’t settle it. The underlying problem is real and documented. For now it’s mostly handled with logs and proxies, not the endpoint tools security teams have spent years learning to trust.

Sources

• AI Needs a Governed Path, Not a Ban, Armor Founder Tells GovWare, PRNewswire, October 14, 2026.
• Claude Cowork architecture overview, Anthropic Help Center.
• Isolation layers: Docker Sandboxes, Docker Docs.
• Nearly Half of Large Enterprises Lack Full Visibility Into AI Use by Employees, According to New Protiviti AI Pulse Survey, Protiviti, May 6, 2026.
• GovWare Conference & Exhibition 2026, Singapore International Cyber Week.

Editorial Disclosure

This article is based on a press release issued by Armor Defense Inc. on October 14, 2026, distributed via PRNewswire. Armor Defense Inc. is a privately held company; no publicly traded security is discussed, and nothing in this article implies that Armor or Sovereign AI is available for retail investment. Next Gen Tech Stocks has not received compensation from Armor Defense Inc., its management, investor relations or public relations representatives, or any third party for this coverage. Statements regarding the capabilities, performance, and availability of Sovereign AI, and regarding demonstrations and briefings at GovWare 2026, are the company’s own and are forward-looking; actual results may differ materially. Survey figures cited are attributed to Protiviti and have not been independently audited by Next Gen Tech Stocks. References to Anthropic, Docker, and Protiviti are for technical context only and do not constitute an endorsement or an investment recommendation.

Coverage on Next Gen Tech Stocks is for informational and educational purposes only and does not constitute professional financial advice. For further details on our publishing standards, advertising relationships, and risk disclaimers, please see our full Terms & Disclaimers Page