Agent Safety Guards

toolCallValidator

Validate tool call arguments for type safety, injection prevention, and tool allowlists.

toolCallValidator({
  action: 'block' | 'warn',
  allowedTools?: string[],    // tool name allowlist
  schemas?: Record<string, object>,  // JSON schemas per tool
})

Detects:

• Missing or invalid tool call arguments
• SQL injection in tool arguments
• Calls to disallowed tools
• Type mismatches against schemas

codeSafety

Detect dangerous code patterns in LLM-generated code.

codeSafety({ action: 'block' | 'warn' })

Detects:

• eval() and Function() calls
• Shell injection (exec, spawn with unsanitized input)
• SQL injection patterns
• Environment variable exposure (process.env)