Every CDN request creates a trust problem between two parties who never meet. A company hands its content to a distribution network it cannot inspect. A stranger in a house somewhere runs a machine that serves that content to browsers. Both sides are reasonable to worry. Evolving Edge, a distributed CDN startup, is publishing a 46-page white paper today called "Two Strangers, One Machine" that lays out how its architecture tries to make as much of that trust as structurally unnecessary as possible.
The question most security teams skip
Traditional security asks how to make a remote node trustworthy. That is the wrong starting point when the hardware belongs to someone else. The operator can read memory, attach a debugger, and capture traffic. Control of the physical machine defeats control of the process running on it every time. Evolving Edge flipped the question to "how little do we have to trust the node at all?" The answer involves their proprietary .ee container format and three encryption levels that determine, above all, who holds the decryption key.
Three levels, one key decision
The levels are straightforward. Level 0 is public content, integrity-checked but not encrypted. Level 1 is the interesting one: the key travels in the URL fragment, the part after the hash symbol that browsers never send to any server. The edge node receives only AES-256-GCM ciphertext. If an operator copies the file off disk, they get random bytes. Even if Evolving Edge's own control plane were compromised, Level 1 content stays unreadable because the company never holds that key either. Level 2 sits between the extremes. The node receives a key for a single authorized request, and the key is released when the response ends. Plaintext exists in operator RAM for the duration of that request.
The distinction between "cannot" and "will not" runs through the entire paper. A承诺 not to read your content is a promise that can be broken, forgotten, or overridden by whoever acquires the company. A cryptographic inability to read it is a property of the system that does not depend on anyone's intentions.
What they say goes wrong
The paper is unusually direct about its own gaps. Section one is called "What We Do Not Claim" and chapter eleven is titled "Limits We State Plainly." Among the items disclosed: Level 2 plaintext lives in operator RAM for the length of a request, and an operator with root access and a debugger can extract it. Evolving Edge recommends Level 1 if operator compromise is in your threat model, even though Level 2 is their more capable mode.
Compression leaks object size. Ciphertext length tracks plaintext content length. In their own tests, they measured differences of 27.5x between same-length files. They list this as a limitation, not a footnote. Crypto-shredding is not yet immediate on running control plane instances, and the paper says so. No SOC 2 Type II or HIPAA certifications exist yet, though readiness work is underway. The paper explicitly states it is not an attestation. And no customer code runs on edge nodes today. AI inference and general edge compute are on the roadmap, but they will require their own threat models and their own papers when there is something to test.
Four controls, not ten
The security team counted their controls honestly and arrived at four. Chapter 9 pairs every security claim with the test that verifies it. Section 9.6 on page 37 lists five things a security team can verify without Evolving Edge's cooperation, starting with opening browser DevTools and watching a Level 1 download arrive as opaque ciphertext, then decrypt in the page.
Chapter 8 was written for edge node operators. It inventories everything the software places on the host machine: a cache of hash-named files the operator cannot read, an encrypted credential file sealed to specific hardware, and a process that serves them. Nothing is hidden, and every claim is written so an operator can check it on their own hardware. The threat model runs both directions, because node operators did not sign up to be trusted either.
The reputation problem
Evolving Edge draws a pointed comparison to platforms like eBay, Uber, and Airbnb. Those systems use reputation mechanisms, reviews and ratings, to make strangers trust each other. That works because there is a window between the transaction and the verdict. Content delivery has no such window. By the time a customer could notice their content was exposed, the exposure is over and no rating reverses it. For confidentiality, reputation is not a weak mechanism. It is an unavailable one. The guarantee has to be a property of the system itself.
The paper is version 1.0. When the limits listed in chapter eleven change, and several are on their list, the paper changes with them and they will document what moved. The web app is live at sqlazy.com, and the edge CDN beta is open for people with idle hardware who want to put it to work.