CVE-2026-33718

OpenHands is software for AI-driven development. Starting in version 1.5.0, a Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels. Version 1.5.0 fixes the issue.

OpenHands is software for AI-driven development. Starting in version 1.5.0, a Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels. Version 1.5.0 fixes the issue.

Summary

A Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels.

Details

Vulnerable Code Path

The vulnerability flows through these files:

1. API Endpoint (openhands/server/routes/files.py:267-277)

@app.get('/git/diff')
async def git_diff(
    path: str,  # <-- User input from HTTP request
    ...
):
    ...
    diff = await call_sync_from_async(runtime.get_git_diff, path, cwd)  # No sanitization

2. Runtime (openhands/runtime/base.py:1231-1233)

def get_git_diff(self, file_path: str, cwd: str) -> dict[str, str]:
    self.git_handler.set_cwd(cwd)
    return self.git_handler.get_git_diff(file_path)  # Passed directly

3. Vulnerable Method (openhands/runtime/utils/git_handler.py:10-12, 134)

# Command template with placeholder
GIT_DIFF_CMD = 'python3 /openhands/code/openhands/runtime/utils/git_diff.py "{file_path}"'

# Line 134 - VULNERABLE: User input directly interpolated
result = self.execute(self.git_diff_cmd.format(file_path=file_path), self.cwd)

4. Shell Execution (openhands/runtime/utils/git_diff.py:25-27)

def run(cmd: str, cwd: str) -> str:
    result = subprocess.run(
        args=cmd,
        shell=True,  # <-- Enables shell metacharacter interpretation
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
        cwd=cwd
    )

Root Cause

The file_path parameter is directly interpolated into a shell command string using Python's .format() method without any sanitization. When this command is executed with shell=True, shell metacharacters like ", ;, and # are interpreted, allowing command injection.

Example:

• Input: test"; id #
• Constructed command: python3 /script.py "test"; id #"
• Shell interprets as two commands: python3 /script.py "test" AND id

Impact

Who is Affected

• All OpenHands deployments exposing the /api/conversations/{id}/git/diff endpoint
• Any authenticated user can exploit this vulnerability,

Attack Capabilities

An attacker can:

1. Execute arbitrary commands on the runtime container as root
2. Read sensitive files including .env, API keys, source code
3. Write arbitrary files to inject malicious code
4. Establish reverse shells for persistent access
5. Potentially escape the container if Docker is misconfigured

Mitigation

Users should update to the latest version of OpenHands that includes the changes from PR #13051. The fix replaces direct shell string formatting with proper argument array handling or rigorous path sanitization to prevent command chaining.

OpenHands is software for AI-driven development. Starting in version 1.5.0, a Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels. Version 1.5.0 fixes the issue.

OpenHands is software for AI-driven development. Starting in version 1.5.0, a Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels. Version 1.5.0 fixes the issue.

Summary

A Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels.

Details

Vulnerable Code Path

The vulnerability flows through these files:

1. API Endpoint (openhands/server/routes/files.py:267-277)

@app.get('/git/diff')
async def git_diff(
    path: str,  # <-- User input from HTTP request
    ...
):
    ...
    diff = await call_sync_from_async(runtime.get_git_diff, path, cwd)  # No sanitization

2. Runtime (openhands/runtime/base.py:1231-1233)

def get_git_diff(self, file_path: str, cwd: str) -> dict[str, str]:
    self.git_handler.set_cwd(cwd)
    return self.git_handler.get_git_diff(file_path)  # Passed directly

3. Vulnerable Method (openhands/runtime/utils/git_handler.py:10-12, 134)

# Command template with placeholder
GIT_DIFF_CMD = 'python3 /openhands/code/openhands/runtime/utils/git_diff.py "{file_path}"'

# Line 134 - VULNERABLE: User input directly interpolated
result = self.execute(self.git_diff_cmd.format(file_path=file_path), self.cwd)

4. Shell Execution (openhands/runtime/utils/git_diff.py:25-27)

def run(cmd: str, cwd: str) -> str:
    result = subprocess.run(
        args=cmd,
        shell=True,  # <-- Enables shell metacharacter interpretation
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
        cwd=cwd
    )

Root Cause

The file_path parameter is directly interpolated into a shell command string using Python's .format() method without any sanitization. When this command is executed with shell=True, shell metacharacters like ", ;, and # are interpreted, allowing command injection.

Example:

• Input: test"; id #
• Constructed command: python3 /script.py "test"; id #"
• Shell interprets as two commands: python3 /script.py "test" AND id

Impact

Who is Affected

• All OpenHands deployments exposing the /api/conversations/{id}/git/diff endpoint
• Any authenticated user can exploit this vulnerability,

Attack Capabilities

An attacker can:

1. Execute arbitrary commands on the runtime container as root
2. Read sensitive files including .env, API keys, source code
3. Write arbitrary files to inject malicious code
4. Establish reverse shells for persistent access
5. Potentially escape the container if Docker is misconfigured

Mitigation

Users should update to the latest version of OpenHands that includes the changes from PR #13051. The fix replaces direct shell string formatting with proper argument array handling or rigorous path sanitization to prevent command chaining.

OpenHands is software for AI-driven development. Starting in version 1.5.0, a Command Injection vulnerability exists in the get_git_diff() method at openhands/runtime/utils/git_handler.py:134. The path parameter from the /api/conversations/{conversation_id}/git/diff API endpoint is passed unsanitized to a shell command, allowing authenticated attackers to execute arbitrary commands in the agent sandbox. The user is already allowed to instruct the agent to execute commands, but this bypasses the normal channels. Version 1.5.0 fixes the issue.