Two critical Atomic Server flaws expose agent private keys
Advisories published on 18 November 2026 describe a stored cross-site scripting bug and a path traversal flaw, both in the Atomic Server download handler.
By The Agentic Times ·

Two critical vulnerabilities in Atomic Server, the Rust-based data server for the Atomic Data specification, were disclosed in GitHub advisories published on 18 November 2026. Both sit in the same /download request handler, and both end in the theft of Ed25519 private keys, according to the advisories. One leads to takeover of a user's account; the other can leak the server root agent's key and hand an attacker full control of the data store.
The first issue, tracked as CVE-2026-69155, concerns how uploaded files are served back to browsers. Files sent to POST /upload are returned by GET /download/{path} using the actix-files library. For HTML and SVG content, the advisory says that library sets a Content-Disposition: inline header, so the browser renders the file rather than saving it. The download route sets no Content Security Policy, no X-Content-Type-Options: nosniff header and does not force an attachment download; a CSP is applied only to the single-page app route. Because uploaded files are served from the same origin as the application interface, a script inside an uploaded HTML file runs with the privileges of the app.
That matters because of where Atomic Server's browser client keeps its secrets. The advisory states that the atomic_session cookie is set through document.cookie and is therefore not marked HttpOnly, and that an agent's Ed25519 private key is stored in same-origin IndexedDB. An attacker who persuades a victim to open an uploaded link can read both and sign commits as that victim indefinitely. The reporter notes this is not limited by the normal ten-second token window, and that if the victim is an admin or root agent the result is server takeover. The behaviour was confirmed at the HTTP layer in a lab using the joepmeneer/atomic-server:develop image, with in-browser key theft described as following from the observed response headers.
The second issue, CVE-2026-69156, is an arbitrary file read. According to the advisory, the commit apply path enforces no allowlist of writable properties, so any authenticated agent can set the internalId string on a resource it owns. The download handler then pushes that value onto a filesystem path and opens it with no sanitisation and no containment check. Setting internalId to a traversal string such as ../../../../etc/passwd, or to an absolute path, lets an attacker read any file the server process can reach. The advisory says only the upload path sanitises filenames, and that the handler never checks the resource is actually a file.
The highest-value target named in the report is the server configuration file holding the root agent's private key. The reproduction steps show reading /atomic-storage/config/config.toml, then using the recovered key to sign a commit as the root agent, giving read, write and delete access to all data. The advisory lists environment data, TLS keys, the sled database and Kubernetes service-account tokens as other readable targets, and says only a standard account is needed.
The suggested fixes differ for each flaw. For the scripting bug, the reporter recommends forcing attachment downloads, adding nosniff and a sandbox CSP, serving user files from a separate origin, and setting HttpOnly and Secure on the session cookie. For the traversal bug, it recommends sanitising and canonicalising the path, confining it to the uploads directory, checking the resource class, and adding a denylist of server-managed properties so clients cannot set internalId.
Sources
- github.comhttps://github.com/advisories/GHSA-x277-3wcg-g9r2
- github.comhttps://github.com/advisories/GHSA-8vc4-8hjq-988p
