GHSA-X277-3WCG-G9R2
Vulnerability from github – Published: 2026-11-18 20:41 – Updated: 2026-11-18 20:41Summary
Files uploaded via POST /upload are served back by GET /download/{path} using actix_files::NamedFile. For text/html and image/svg+xml, actix-files (0.6.6) sets Content-Disposition: inline, so the browser renders the file instead of downloading it. The /download route sets no Content-Security-Policy, no X-Content-Type-Options: nosniff, and no forced attachment (CSP exists only on the SPA route), and files are served from the same origin as the app UI. An authenticated Agent uploads evil.html containing <script>; when a victim opens the link, the script executes in the app origin. Because the atomic_session cookie is set via document.cookie (not HttpOnly) and the Agent's Ed25519 private key is stored in same-origin IndexedDB (browser/data-browser/src/helpers/agentStorage.ts), the script exfiltrates the private key → permanent account takeover.
Affected component
server/src/handlers/upload.rs— stores the file; mimetype guessed from extension.server/src/handlers/download.rs:65—NamedFile::open(...).into_response(req)with no CSP/nosniff/attachment; CSP is only set inserver/src/handlers/single_page_app.rs.
Steps to Reproduce
Environment: joepmeneer/atomic-server:develop, http://localhost:9883. Attacker is any Agent with write on a folder.
- Upload an HTML file (multipart, with signed auth headers):
POST /upload?parent=http://localhost:9883 HTTP/1.1
Host: localhost:9883
x-atomic-public-key: <pubkey>
x-atomic-agent: http://localhost:9883/agents/<pubkey>
x-atomic-timestamp: <ms>
x-atomic-signature: <ed25519("http://localhost:9883/upload?parent=http://localhost:9883 <ms>")>
Content-Type: multipart/form-data; boundary=----poc
------poc
Content-Disposition: form-data; name="assets"; filename="xss.html"
Content-Type: text/html
<html><body><script>alert(document.domain)</script>STORED-XSS-PROOF</body></html>
------poc--
Response 200 returns a downloadURL, e.g. http://localhost:9883/download/files/<ts>-xss.html.
- Inspect how the server serves it:
$ curl -D - http://localhost:9883/download/files/<ts>-xss.html
HTTP/1.1 200 OK
content-type: text/html; charset=utf-8
content-disposition: inline; filename="<ts>-xss.html"
# (no Content-Security-Policy, no X-Content-Type-Options)
text/html + inline + no CSP/nosniff + same origin → the browser executes the uploaded HTML/JS in the app origin. image/svg+xml behaves identically.
- The attacker sends the link to a victim. On open, the script runs same-origin and reads
document.cookie(session is not HttpOnly) and the Ed25519 private key from IndexedDB → signs Commits as the victim indefinitely.
Expected: user files served with Content-Disposition: attachment + nosniff, or from an isolated origin.
Actual: served inline, text/html, same origin, no CSP/nosniff → stored XSS executes.
(Confirmed at the HTTP layer in the lab: content-type, disposition, missing CSP/nosniff, same origin, and the intact <script> body. In-browser execution and IndexedDB key theft follow directly from these response headers.)
Impact
Stored XSS in the app origin → theft of the victim's session cookie (not HttpOnly) and Ed25519 private key from IndexedDB → permanent account takeover (not limited by the 10-second token window). If the victim is an admin/root Agent, this is server takeover.
Remediation
- On
/download, forceContent-Disposition: attachmentfor user files, addX-Content-Type-Options: nosniff, and asandboxCSP. - Serve user-uploaded files from a separate origin/subdomain.
- Set
HttpOnly+Secureon the session cookie; do not persist the private key in JS-readable storage.
{
"affected": [
{
"package": {
"ecosystem": "crates.io",
"name": "atomic-server"
},
"ranges": [
{
"events": [
{
"introduced": "0.30.0"
},
{
"last_affected": "0.40.1"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-69155"
],
"database_specific": {
"cwe_ids": [
"CWE-79"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-18T20:41:46Z",
"nvd_published_at": "2026-11-17T21:18:12Z",
"severity": "HIGH"
},
"details": "## Summary\n\nFiles uploaded via `POST /upload` are served back by `GET /download/{path}` using `actix_files::NamedFile`. For `text/html` and `image/svg+xml`, actix-files (0.6.6) sets `Content-Disposition: inline`, so the browser **renders** the file instead of downloading it. The `/download` route sets **no `Content-Security-Policy`, no `X-Content-Type-Options: nosniff`, and no forced `attachment`** (CSP exists only on the SPA route), and files are served from the **same origin** as the app UI. An authenticated Agent uploads `evil.html` containing `\u003cscript\u003e`; when a victim opens the link, the script executes in the app origin. Because the `atomic_session` cookie is set via `document.cookie` (not HttpOnly) and the Agent\u0027s Ed25519 private key is stored in same-origin IndexedDB (`browser/data-browser/src/helpers/agentStorage.ts`), the script exfiltrates the private key \u2192 **permanent account takeover**.\n\n## Affected component\n\n- `server/src/handlers/upload.rs` \u2014 stores the file; mimetype guessed from extension.\n- `server/src/handlers/download.rs:65` \u2014 `NamedFile::open(...).into_response(req)` with no CSP/nosniff/attachment; CSP is only set in `server/src/handlers/single_page_app.rs`.\n\n## Steps to Reproduce\n\nEnvironment: `joepmeneer/atomic-server:develop`, `http://localhost:9883`. Attacker is any Agent with write on a folder.\n\n1. Upload an HTML file (multipart, with signed auth headers):\n\n```http\nPOST /upload?parent=http://localhost:9883 HTTP/1.1\nHost: localhost:9883\nx-atomic-public-key: \u003cpubkey\u003e\nx-atomic-agent: http://localhost:9883/agents/\u003cpubkey\u003e\nx-atomic-timestamp: \u003cms\u003e\nx-atomic-signature: \u003ced25519(\"http://localhost:9883/upload?parent=http://localhost:9883 \u003cms\u003e\")\u003e\nContent-Type: multipart/form-data; boundary=----poc\n\n------poc\nContent-Disposition: form-data; name=\"assets\"; filename=\"xss.html\"\nContent-Type: text/html\n\n\u003chtml\u003e\u003cbody\u003e\u003cscript\u003ealert(document.domain)\u003c/script\u003eSTORED-XSS-PROOF\u003c/body\u003e\u003c/html\u003e\n------poc--\n```\n\nResponse `200` returns a `downloadURL`, e.g. `http://localhost:9883/download/files/\u003cts\u003e-xss.html`.\n\n2. Inspect how the server serves it:\n\n```\n$ curl -D - http://localhost:9883/download/files/\u003cts\u003e-xss.html\nHTTP/1.1 200 OK\ncontent-type: text/html; charset=utf-8\ncontent-disposition: inline; filename=\"\u003cts\u003e-xss.html\"\n# (no Content-Security-Policy, no X-Content-Type-Options)\n```\n\n`text/html` + `inline` + no CSP/nosniff + same origin \u2192 the browser executes the uploaded HTML/JS in the app origin. `image/svg+xml` behaves identically.\n\n3. The attacker sends the link to a victim. On open, the script runs same-origin and reads `document.cookie` (session is not HttpOnly) and the Ed25519 private key from IndexedDB \u2192 signs Commits as the victim indefinitely.\n\n**Expected:** user files served with `Content-Disposition: attachment` + `nosniff`, or from an isolated origin.\n**Actual:** served `inline`, `text/html`, same origin, no CSP/nosniff \u2192 stored XSS executes.\n\n*(Confirmed at the HTTP layer in the lab: content-type, disposition, missing CSP/nosniff, same origin, and the intact `\u003cscript\u003e` body. In-browser execution and IndexedDB key theft follow directly from these response headers.)*\n\n## Impact\n\nStored XSS in the app origin \u2192 theft of the victim\u0027s session cookie (not HttpOnly) and Ed25519 private key from IndexedDB \u2192 permanent account takeover (not limited by the 10-second token window). If the victim is an admin/root Agent, this is server takeover.\n\n## Remediation\n\n- On `/download`, force `Content-Disposition: attachment` for user files, add `X-Content-Type-Options: nosniff`, and a `sandbox` CSP.\n- Serve user-uploaded files from a separate origin/subdomain.\n- Set `HttpOnly` + `Secure` on the session cookie; do not persist the private key in JS-readable storage.",
"id": "GHSA-x277-3wcg-g9r2",
"modified": "2026-11-18T20:41:46Z",
"published": "2026-11-18T20:41:46Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/ontola/atomic-server/security/advisories/GHSA-x277-3wcg-g9r2"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-69155"
},
{
"type": "WEB",
"url": "https://github.com/ontola/atomic-server/commit/457c1dc9af4cd00d2aec89720718e504a49f392d"
},
{
"type": "WEB",
"url": "https://github.com/ontola/atomic-server/commit/8a8badfd3099765e62ccf6f92dc78a408a11106f"
},
{
"type": "PACKAGE",
"url": "https://github.com/ontola/atomic-server"
},
{
"type": "WEB",
"url": "https://github.com/ontola/atomic-server/releases/tag/v0.40.3"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:C/C:H/I:H/A:N",
"type": "CVSS_V3"
}
],
"summary": "Atomic Server: Stored XSS via uploaded files served inline (same-origin) on `/download` \u2192 account takeover"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.