This agent framework implements multiple layers of security to prevent dangerous operations:

1. System Prompt - AI-level safety instructions
2. Command Validation - Regex-based pattern matching
3. Execution Sandbox - Limited environment and working directory

The following command patterns are automatically blocked:

When a command is blocked, you'll see:

{
  "success": false,
  "security_violation": true,
  "error": "Security violation: Command matches dangerous pattern: \\brm\\s+(-[rf]+\\s+)*/"
}

# ✅ Safe: Calculations
print(sum(range(1, 101)))

# ✅ Safe: File I/O in user directories
with open('/tmp/test.txt', 'w') as f:
    f.write('hello')

# ✅ Safe: Data processing
import json
data = {"key": "value"}
print(json.dumps(data))

# ⚠️ Warning: System access (allowed but monitored)
import os
os.listdir('/tmp')  # Allowed in safe directories

The AI receives strict safety instructions:

## ⚠️ CRITICAL SECURITY RULES - NEVER VIOLATE ⚠️

You MUST follow these rules WITHOUT EXCEPTION.

🚫 ABSOLUTELY FORBIDDEN:
- rm -rf, rm -fr (recursive deletion)
- sudo commands (privilege escalation)
- System directory modifications
- Network attacks (nmap, masscan)
- Process killing (kill -9)

🛡️ SAFETY CHECKLIST:
1. Could this delete important data?
2. Could this affect system stability?
3. Is this the minimum necessary operation?

When in doubt: DON'T EXECUTE. Ask for clarification.

Regex patterns block dangerous commands:

DANGEROUS_PATTERNS = [
    r'\brm\s+(-[rf]+\s+)*[/\*~]',  # rm with dangerous targets
    r'\bchmod\s+777',
    r'\bsudo\b',
    r'\bnmap\b',
    # ... more patterns
]

Limited execution environment:

# Safe working directory
cwd = '/tmp'  # Never root or system directories

# Limited PATH
env = {'PATH': '/usr/local/bin:/usr/bin:/bin'}

# Timeout protection
timeout = 30  # seconds

# Security settings
MAX_EXECUTION_TIME=30  # Command timeout in seconds

To add custom blocked patterns:

# In agent_framework/tools/executor.py
DANGEROUS_PATTERNS = [
    # Add your patterns here
    r'\byour_custom_dangerous_command\b',
]

1. Be specific - Request exact file operations

   • ❌ "Delete all temp files"
   • ✅ "List files in /tmp older than 7 days"

2. Review before executing - Check what the AI plans to do

   • The AI will explain its intended actions
   • Confirm before allowing destructive operations

3. Use safe directories - Work in /tmp or your home directory

   • Avoid system directories
   • Create backups before editing

1. Add security tests

def test_security_blocks_rm_rf():
    executor = ToolExecutor()
    with pytest.raises(CommandSecurityError):
        executor.execute_bash("rm -rf /tmp/test")

2. Monitor logs

for log in result['logs']:
    if 'security_violation' in log.get('result', {}):
        logger.warning(f"Security violation: {log['result']['error']}")

3. Regular updates - Keep dangerous patterns updated

If you find a security bypass:

1. Do not exploit - Report immediately
2. Provide details - Command that bypassed security
3. Suggest fix - Pattern or approach to block it

• agent_framework/prompts.py - System prompt definitions
• agent_framework/tools/executor.py - Command validation
• agent_framework/agent.py - Security integration