AI Agent oddiy chatbot emas u filesystem, terminal, API, browser, database va boshqa tool’lar bilan ishlashi mumkin. Shuning uchun production yoki lab muhitida ishga tushirishdan oldin quyidagilarni tekshiring:
1. Least Privilege
Agentga faqat vazifasi uchun zarur bo‘lgan permissions bering.
root, Administrator yoki ortiqcha IAM privileges bilan ishlatmang.2. Sandbox Isolation
Code execution va file processing’ni alohida Docker container, VM, WSL yoki sandbox ichida bajaring. Host filesystem’ga unrestricted access bermang.
3. Secrets Management
API key, token, password, .env, SSH key va credential’larni prompt/context ichiga joylashtirmang. Secrets’ni alohida Secrets Manager orqali boshqaring.4. Tool Permissions
shell, filesystem, browser, email, database va boshqa tool’larni default-deny tamoyili bilan ulang. Keraksiz tool keraksiz attack surface degani.5. Network Egress Control
Agent istalgan domain yoki IP’ga chiqmasligi kerak. allowlist, proxy yoki firewall orqali outbound traffic’ni cheklang.
6. Prompt Injection Defense
Web page, PDF, email, source code va RAG orqali kelgan content’ni untrusted input deb qabul qiling. External instruction’lar system prompt yoki security policy’ni override qilmasligi kerak.
7. Command Execution Guardrails
rm, curl | bash, PowerShell, package installation, process execution kabi high-risk action’lar uchun validation yoki human approval qo‘llang.8. Data Access Scope
Agentga butun disk, database yoki shared folder’ni bermang. Faqat kerakli directory, table, collection yoki dataset bilan ishlashiga ruxsat bering.
9. Logging & Audit Trail
Prompt, tool call, executed command, file access, network request va error’larni audit log qiling. Incident yuz bersa, agent nima qilganini reconstruct qilish mumkin bo‘lishi kerak.
10. Kill Switch & Monitoring
Agent behaviour’ini kuzating va abnormal activity aniqlansa uni darhol to‘xtatadigan kill switch, rate limit, timeout va resource limit o‘rnating.
Asosiy qoida:
AI Agent’ni “aqlli assistant” emas, untrusted code’ni execute qilishi mumkin bo‘lgan privileged automation system sifatida himoyalang.
Local model ishlatayotganingiz data avtomatik ravishda xavfsiz degani emas - model local bo‘lishi mumkin, attack surface esa local emas.
🇺🇿 CYBER-BRO - Xavfsiz kiber sarhadlar sari!
X - Telegram - Linkedin - Youtube