Safe Tool Executor

Enforces safe tool execution with tier-based access control and human approval for dangerous operations.

Problem: Unsafe Tool Use

SYMPTOMS:
├── Over-privileged tools
├── Destructive action undetected
├── Deletion of important data
└── Insufficient watchdog severity

Solutions Implemented

1. Least Privilege Principle

class SafeToolExecutor:
    tool_tiers = {
        'READ_ONLY': ['ls', 'cat', 'head', 'tail', 'grep'],
        'WRITE': ['write', 'edit', 'mkdir', 'touch'],
        'DELETE': ['rm', 'rmdir', 'unlink']
    }
    
    required_approval = ['DELETE', 'DROP', 'TRUNCATE', 'FORMAT']

2. Approval Workflow

def execute_dangerous_tool(action, tool, args):
    if tool in required_approval:
        request_human_approval(f"DANGER: {tool} {args}")
        wait_for_approval(timeout=60)
        if not approved:
            return {'status': 'BLOCKED', 'reason': 'No approval'}
    return execute_tool(tool, args)

3. Tier Separation

def validate_tool_access(tool, operation):
    # READ_ONLY tools cannot write
    if operation == 'WRITE' and tool in tool_tiers['READ_ONLY']:
        return False
    # DELETE tools require approval by default
    if tool in tool_tiers['DELETE']:
        return require_approval(tool)
    return True

Watchdogs

Usage

from safe_tool_executor import SafeToolExecutor

executor = SafeToolExecutor()

# READ_ONLY tool - OK directly
result = executor.execute('cat', '/etc/passwd')

# WRITE tool - Warning
result = executor.execute('write', '/project/config.py')

# DELETE tool - BLOCKED without approval
result = executor.execute('rm', '/important/file.txt')
# → BLOCKED: requires human approval

Dangerous Patterns (Blocked)

Prerequisites

Files

safe-tool-executor/
├── SKILL.md
├── scripts/
│   ├── safe_tool_executor.py
│   ├── main.py
│   └── utils.py
├── data/
├── models/
└── tests/