GHSA-79WM-X847-7CVG
Vulnerability from github – Published: 2026-09-03 19:50 – Updated: 2026-09-03 19:50Summary
npx claude-code-templates --studio launches "Claude Code Studio", an Express HTTP server (cli-tool/src/sandbox-server.js, default port 3444) that binds to all interfaces (0.0.0.0), sets Access-Control-Allow-Origin: *, and requires no authentication. Two POST endpoints pass attacker-controlled request-body fields into child_process.spawn(..., { shell: true }). Because shell: true makes Node join the argv array into a single sh -c string, the fields are parsed by the shell and metacharacters execute. Any unauthenticated attacker who can reach the port — a malicious web page the developer visits, or anyone on the same LAN — can execute arbitrary OS commands on the developer's machine.
Details
In cli-tool/src/sandbox-server.js:
app.listen(PORT, ...)is called with no host argument, so the server listens on0.0.0.0/::(reachable from the LAN, not just localhost).- The CORS middleware sends
Access-Control-Allow-Origin: *and answers the preflightOPTIONSfor any origin, so a browser will deliver cross-origin POSTs to it. - There is no authentication on any endpoint.
The vulnerable sinks:
-
POST /api/execute— thepromptbody field flows intoexecuteLocalTask(): ```js const child = spawn('claude', [finalPrompt], { / ... / shell: true }); The only validation on prompt is a length check (>= 10 chars). With shell: true, finalPrompt is interpreted by the shell. -
POST /api/install-agent — the agentName body field: const child = spawn('npx', ['claude-code-templates@latest', '--agent', agentName, '--yes'], { / ... / shell: true });
- agentName is used unvalidated. (The same unsafe pattern is also reachable through /api/execute's agent field via checkAndInstallAgent().)
Root cause: spawn(cmd, argsArray, { shell: true }) does not keep argsArray as separate argv entries — Node builds cmd + ' ' + argsArray.join(' ') and runs it via sh -c, so every element is subject to shell parsing.
PoC
Victim
npx claude-code-templates --studio # server on 0.0.0.0:3444
Attacker (another LAN host, or a malicious web page fetch(), or locally)
curl -s -X POST http://127.0.0.1:3444/api/execute \ -H 'Content-Type: application/json' \ --data '{"prompt":"aaaaaaaaaa; touch /tmp/CCT_RCE_PROOF","mode":"local"}'
curl -s -X POST http://127.0.0.1:3444/api/install-agent \ -H 'Content-Type: application/json' \ --data '{"agentName":"x; touch /tmp/CCT_AGENT_PROOF #"}'
ls -la /tmp/CCT_RCE_PROOF /tmp/CCT_AGENT_PROOF # both created => injected commands ran The aaaaaaaaaa padding satisfies the 10-char minimum, then ; (or $(...), or backticks) starts the injected command. claude/npx do not even need to be installed — the injected segment runs regardless.
Confirmed at runtime on v1.28.13 (Node 22, Linux): both marker files were created, the server listened on :3444, and an OPTIONS preflight from Origin: https://evil.example returned 200 with Access-Control-Allow-Origin: .
Impact
Unauthenticated remote code execution (CWE-78) on any machine running --studio. Two reachability paths: - Drive-by: a developer running --studio who visits an attacker-controlled web page — the page's cross-origin fetch() (Content-Type application/json) passes the wildcard CORS preflight and delivers the POST, achieving RCE with no other interaction. - LAN: because the server binds 0.0.0.0, anyone on the same network (office, co-working space, public Wi-Fi) can hit port 3444 directly.
Impact is full compromise of the developer's user account (arbitrary command execution with the developer's privileges): source code, SSH keys, cloud credentials, and .env secrets.
Suggested fix
- Remove shell: true from all three spawns so arguments stay discrete argv entries (kills the injection).
- Validate agentName against a strict allowlist (^[A-Za-z0-9._/-]+$).
- Bind to loopback only (app.listen(PORT, '127.0.0.1', ...)).
- Replace the wildcard CORS with a same-origin allowlist and reject other origins.
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 1.29.2"
},
"package": {
"ecosystem": "npm",
"name": "claude-code-templates"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "1.29.4"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-73222"
],
"database_specific": {
"cwe_ids": [
"CWE-306",
"CWE-352",
"CWE-78"
],
"github_reviewed": true,
"github_reviewed_at": "2026-09-03T19:50:47Z",
"nvd_published_at": "2026-08-11T19:18:51Z",
"severity": "HIGH"
},
"details": "### Summary\n`npx claude-code-templates --studio` launches \"Claude Code Studio\", an Express HTTP server (`cli-tool/src/sandbox-server.js`, default port 3444) that binds to **all interfaces** (`0.0.0.0`), sets `Access-Control-Allow-Origin: *`, and requires **no authentication**. Two POST endpoints pass attacker-controlled request-body fields into `child_process.spawn(..., { shell: true })`. Because `shell: true` makes Node join the argv array into a single `sh -c` string, the fields are parsed by the shell and metacharacters execute. Any unauthenticated attacker who can reach the port \u2014 a malicious web page the developer visits, or anyone on the same LAN \u2014 can execute arbitrary OS commands on the developer\u0027s machine.\n\n### Details\nIn `cli-tool/src/sandbox-server.js`:\n\n- `app.listen(PORT, ...)` is called with no host argument, so the server listens on `0.0.0.0` / `::` (reachable from the LAN, not just localhost).\n- The CORS middleware sends `Access-Control-Allow-Origin: *` and answers the preflight `OPTIONS` for any origin, so a browser will deliver cross-origin POSTs to it.\n- There is no authentication on any endpoint.\n\nThe vulnerable sinks:\n\n1. `POST /api/execute` \u2014 the `prompt` body field flows into `executeLocalTask()`:\n ```js\n const child = spawn(\u0027claude\u0027, [finalPrompt], { /* ... */ shell: true });\n The only validation on prompt is a length check (\u003e= 10 chars). With shell: true, finalPrompt is interpreted by the shell.\n\n2. POST /api/install-agent \u2014 the agentName body field:\nconst child = spawn(\u0027npx\u0027, [\u0027claude-code-templates@latest\u0027, \u0027--agent\u0027, agentName, \u0027--yes\u0027], { /* ... */ shell: true });\n2. agentName is used unvalidated. (The same unsafe pattern is also reachable through /api/execute\u0027s agent field via checkAndInstallAgent().)\n\nRoot cause: spawn(cmd, argsArray, { shell: true }) does not keep argsArray as separate argv entries \u2014 Node builds cmd + \u0027 \u0027 + argsArray.join(\u0027 \u0027) and runs it via sh -c, so every element is subject to shell parsing.\n\nPoC\n\n# Victim\nnpx claude-code-templates --studio # server on 0.0.0.0:3444\n\n# Attacker (another LAN host, or a malicious web page fetch(), or locally)\ncurl -s -X POST http://127.0.0.1:3444/api/execute \\\n -H \u0027Content-Type: application/json\u0027 \\\n --data \u0027{\"prompt\":\"aaaaaaaaaa; touch /tmp/CCT_RCE_PROOF\",\"mode\":\"local\"}\u0027\n\ncurl -s -X POST http://127.0.0.1:3444/api/install-agent \\\n -H \u0027Content-Type: application/json\u0027 \\\n --data \u0027{\"agentName\":\"x; touch /tmp/CCT_AGENT_PROOF #\"}\u0027\n\nls -la /tmp/CCT_RCE_PROOF /tmp/CCT_AGENT_PROOF # both created =\u003e injected commands ran\nThe aaaaaaaaaa padding satisfies the 10-char minimum, then ; (or $(...), or backticks) starts the injected command. claude/npx do not even need to be installed \u2014 the injected segment runs regardless.\n\nConfirmed at runtime on v1.28.13 (Node 22, Linux): both marker files were created, the server listened on *:3444, and an OPTIONS preflight from Origin: https://evil.example returned 200 with Access-Control-Allow-Origin: *.\n\nImpact\n\nUnauthenticated remote code execution (CWE-78) on any machine running --studio. Two reachability paths:\n- Drive-by: a developer running --studio who visits an attacker-controlled web page \u2014 the page\u0027s cross-origin fetch() (Content-Type application/json) passes the wildcard CORS preflight and delivers the POST, achieving RCE with no other interaction.\n- LAN: because the server binds 0.0.0.0, anyone on the same network (office, co-working space, public Wi-Fi) can hit port 3444 directly.\n\nImpact is full compromise of the developer\u0027s user account (arbitrary command execution with the developer\u0027s privileges): source code, SSH keys, cloud credentials, and .env secrets.\n\nSuggested fix\n\n- Remove shell: true from all three spawns so arguments stay discrete argv entries (kills the injection).\n- Validate agentName against a strict allowlist (^[A-Za-z0-9._/-]+$).\n- Bind to loopback only (app.listen(PORT, \u0027127.0.0.1\u0027, ...)).\n- Replace the wildcard CORS with a same-origin allowlist and reject other origins.",
"id": "GHSA-79wm-x847-7cvg",
"modified": "2026-09-03T19:50:47Z",
"published": "2026-09-03T19:50:47Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/davila7/claude-code-templates/security/advisories/GHSA-79wm-x847-7cvg"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-73222"
},
{
"type": "WEB",
"url": "https://github.com/davila7/claude-code-templates/commit/bc4618b07232633c1c0aac12a43e436268d31783"
},
{
"type": "PACKAGE",
"url": "https://github.com/davila7/claude-code-templates"
},
{
"type": "WEB",
"url": "https://github.com/davila7/claude-code-templates/blob/main/CHANGELOG.md"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H",
"type": "CVSS_V3"
}
],
"summary": "Claude Code Templates: Unauthenticated OS command injection (RCE) in Claude Code Studio server (--studio)"
}
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.