GHSA-VJV9-7M7J-H833

The published npm package praisonai exports SandboxExecutor, CommandValidator, and sandboxExec as "safe command execution with restrictions." When allowedCommands is configured, CommandValidator checks only the first whitespace-delimited token of the command string. SandboxExecutor then passes the entire original string to spawn("sh", ["-c", command]).

With a policy that allows only echo, this direct command is correctly rejected:

cat /tmp/marker

but this chained command is accepted and executed:

echo allowed; cat /tmp/marker

The shell executes cat even though cat is not allowlisted. This bypasses the command allowlist and can execute arbitrary shell commands with the PraisonAI process privileges when an application, CLI workflow, or agent pipeline exposes sandbox command execution to lower-trust users, prompts, or model output.

The PoV is deterministic and local-only. It creates and reads only a temporary marker file.

Technical Details

In src/praisonai-ts/src/cli/features/sandbox-executor.ts, CommandValidator.validate() normalizes the command and authorizes only the first whitespace token:

const normalized = command.toLowerCase().trim();

if (this.allowedCommands) {
  const baseCmd = normalized.split(/\s+/)[0];
  if (!this.allowedCommands.includes(baseCmd)) {
    return { valid: false, reason: `Command '${baseCmd}' not in allowlist` };
  }
}

The denylist does not generally reject shell separators. It blocks a few specific patterns such as ; rm, but not ; cat, &&, ||, backticks, $(), or newline as a general policy boundary.

SandboxExecutor.spawn() then executes the unmodified command string through a shell:

const proc = spawn('sh', ['-c', command], {
  cwd: this.config.cwd,
  env,
  timeout: this.config.timeout,
  stdio: ['pipe', 'pipe', 'pipe']
});

That creates a mismatch: the allowlist authorizes one command token, but the shell interprets the whole string as a script.

The published npm:[email protected] dist files preserve the same behavior:

• dist/cli/features/sandbox-executor.js checks only baseCmd.
• dist/cli/features/sandbox-executor.js later invokes spawn("sh", ["-c", command]).
• dist/index.js exports SandboxExecutor, CommandValidator, and sandboxExec.

Why This Is Not Intended Behavior

PraisonAI's sandbox docs describe sandbox execution as a security feature for AI-generated commands, with command validation, resource limits, path restrictions, network isolation, and execution isolation. The TypeScript source also describes this component as "Safe command execution with restrictions."

With allowedCommands: ["echo"], PraisonAI correctly rejects cat <marker> when submitted directly. That proves the intended policy is to block non-allowlisted executables. The same policy allowing echo allowed; cat <marker> is therefore an authorization bypass, not merely a permissive configuration.

PoV

Run from a local reproduction checkout:

node poc/pov_poc.js 1.7.1

Expected output includes:

{
  "version": "1.7.1",
  "package": "npm:praisonai",
  "allowedCommands": ["echo"],
  "controls": {
    "directCatRejected": true,
    "benignEchoAllowed": true,
    "patchedControlRejectsChainedShell": true
  },
  "observed": {
    "directPolicy": {
      "allowed": false,
      "reason": "Command 'cat' not in allowlist"
    },
    "benignPolicy": {
      "allowed": true
    },
    "chainedPolicy": {
      "allowed": true
    },
    "chainedRun": {
      "success": true,
      "stdout": "allowed\npoc.7.1",
      "stderr": "",
      "exitCode": 0
    },
    "patchedControl": {
      "benign": {
        "allowed": true
      },
      "direct": {
        "allowed": false,
        "reason": "Command 'cat' not in allowlist"
      },
      "chained": {
        "allowed": false,
        "reason": "shell metacharacter rejected before execution"
      }
    }
  },
  "vulnerable": true
}

Interpretation:

• Direct cat <marker> is rejected by the allowlist.
• Benign echo allowed is accepted.
• echo allowed; cat <marker> is accepted by the same allowlist and executes the non-allowlisted cat.
• A patched-control validator that rejects shell metacharacters before execution blocks the chained command while still allowing benign echo.

The PoV installs npm:[email protected] into a temporary project, creates a temporary marker file, and reads only that file. It does not contact any live service or execute destructive commands.

PoC

The PoV section above contains the local reproduction command, input, and decisive output.

Severity

Suggested severity: High.

Rationale:

• AV: common deployment pattern is an application exposing agent prompts or command automation over a network.
• AC: attacker only needs to induce or submit a command string that starts with an allowed command.
• PR: conservative base score assumes the attacker can submit prompts or command requests to the application.
• UI: no operator action is needed once the command reaches the executor.
• S: impact is in the PraisonAI-hosting process.
• C/I/A: arbitrary shell commands can affect confidentiality, integrity, and availability depending on process privileges.

If maintainers score only local CLI use, AV:L may be reasonable. If they score public unauthenticated prompt or command endpoints built on this API, PR:N may be reasonable.

Affected Package/Versions

• Repository: MervinPraison/PraisonAI
• Package: npm:praisonai
• Component: TypeScript CLI feature SandboxExecutor
• Current head validated: 1ad58ca02975ff1398efeda694ea2ab78f20cf3e
• Current tag validated: v4.6.58
• Latest npm package validated: 1.7.1

Suggested affected range:

npm:praisonai >= 1.2.3, <= 1.7.1

Selected version sweep:

• 1.0.0: package main cannot be required in the selected test environment.
• 1.2.0, 1.2.1, 1.2.2: SandboxExecutor is not exported.
• 1.2.3: vulnerable.
• 1.2.4: vulnerable.
• 1.3.0: vulnerable.
• 1.3.6: vulnerable.
• 1.4.0: vulnerable.
• 1.5.0: vulnerable.
• 1.5.4: vulnerable.
• 1.6.0: vulnerable.
• 1.7.0: vulnerable.
• 1.7.1: vulnerable.

Advisory History

This is distinct from known and previously submitted PraisonAI issues:

• GHSA-r4f2-3m54-pp7q covers PyPI SubprocessSandbox shell=True and blocklist bypass.
• GHSA-2763-cj5r-c79m covers PyPI praisonai OS command injection.
• GHSA-v7px-3835-7gjx covers PyPI memory/hooks.py shell injection.
• GHSA-4wr3-f4p3-5wjh covers Python agent tool approval allow-list manipulation.
• GHSA-4mr5-g6f9-cfrh covers PyPI/Python execute_code sandbox escape.
• GHSA-9qhq-v63v-fv3j covers an incomplete fix for a Python command injection.
• GHSA-vmmj-pfw7-fjwp covers npm codeMode host-process new Function sandbox escape.

No visible local or GitHub advisory covers npm TypeScript SandboxExecutor, CommandValidator, allowedCommands, or the first-token allowlist followed by sh -c shell-chaining root cause.

If lower-trust users, prompts, or model output can influence a command string sent to SandboxExecutor or sandboxExec, allowedCommands does not enforce the intended command boundary. An attacker can append arbitrary shell commands after an allowed first token and run them with the privileges of the PraisonAI process.

Concrete consequences depend on the hosting application and configured process privileges, but can include reading or modifying files, invoking local tools, using available credentials, or causing denial of service.

This report does not claim that npm PraisonAI exposes this as a default network service. It is a library-level sandbox/allowlist bypass in an exported TypeScript API that is explicitly designed for safe command execution.

Untrusted input reaches a shell command, allowing arbitrary commands to run on the host. Typical impact: code execution in the application's environment.

GHSA-VJV9-7M7J-H833 has a CVSS score of 8.8 (High). The vector is network-reachable, low privileges required, and no user interaction. A CVSS score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether this affects your application depends on whether the vulnerable code is present and reachable in your environment. A fixed version is available (1.7.2); upgrading removes the vulnerable code path.

Kodem intelligence

Avoid passing policy-checked user strings to a shell.

Recommended:

1. Require callers to pass { command, args }, or parse command strings into argv with a shell-aware parser.
2. Execute with spawn(command, args, { shell: false }) / execFile() instead of sh -c.
3. Apply allowedCommands to the exact executable after normalization.
4. Reject shell metacharacters (;, &&, ||, |, backticks, $(), newline, redirects) when a shell string API must be kept for compatibility.
5. Add regression tests proving allowedCommands: ["echo"] allows echo ok but rejects cat marker, echo ok; cat marker, echo ok && cat marker, and echo ok | cat marker.

1. What is GHSA-VJV9-7M7J-H833? GHSA-VJV9-7M7J-H833 is a high-severity OS command injection vulnerability in praisonai (npm), affecting versions >= 1.2.3, <= 1.7.1. It is fixed in 1.7.2. Untrusted input reaches a shell command, allowing arbitrary commands to run on the host.
2. How severe is GHSA-VJV9-7M7J-H833? GHSA-VJV9-7M7J-H833 has a CVSS score of 8.8 (High). This score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether it represents real risk in your environment depends on whether the vulnerable code is present and reachable.
3. Which versions of praisonai are affected by GHSA-VJV9-7M7J-H833? praisonai (npm) versions >= 1.2.3, <= 1.7.1 is affected.
4. Is there a fix for GHSA-VJV9-7M7J-H833? Yes. GHSA-VJV9-7M7J-H833 is fixed in 1.7.2. Upgrade to this version or later.
5. Is GHSA-VJV9-7M7J-H833 exploitable, and should I be worried? Whether GHSA-VJV9-7M7J-H833 is exploitable in your environment depends on whether the vulnerable code is present and reachable. A CVSS score is a worst-case rating; it does not account for your specific deployment, configuration, or usage patterns. Kodem, an Intelligent Application Security platform, uses runtime intelligence to show which vulnerabilities actually execute in production, so you can focus on the ones that represent real risk. Get a demo
6. What actually determines whether GHSA-VJV9-7M7J-H833 is exploitable, and how bad it is? Exploitability and impact are not fixed properties of a CVE. They depend on runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A high CVSS score on a dependency that never runs is not the same as real risk. Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter.
7. How do I fix GHSA-VJV9-7M7J-H833? Upgrade praisonai to 1.7.2 or later.