I built an autonomous multi-agent AI pentester — and why it's not another GPT wrapper

Описание и применение

OIHK – это автономный, многоагентный движок для проведения пентестов, который требует реального выполнения инструментов и независимой валидации, чтобы подтвердить уязвимости, тем самым устраняя ложные срабатывания и повышая безопасность.

Задача

Traditional AI pentester projects rely on a single LLM in a loop, producing confident but often false vulnerability reports. There is a need for a system where findings are earned through verifiable evidence and validated by independent agents.

Результат

OIHK delivers verifiable penetration testing by requiring real tool execution and independent validation, reducing false positives and enhancing safety through code‑based guardrails.

Рабочий процесс

• 1. Define exact target scope and resolve hosts once.
• 2. Root planner creates a versioned scan plan and assigns tasks to specialist agents.
• 3. Recon agent gathers reconnaissance data and records evidence.
• 4. Discovery agent performs automated scans, executing tools in a hardened sandbox and logging execution records.
• 5. Validation agent independently verifies findings using separate tool executions and attaches validation evidence.
• 6. Reporting agent compiles findings into a report, referencing the evidence ledger.
• 7. The run is closed only when all critical tasks are completed and validated.
• 8. Optionally run the evaluation mode to benchmark the model against 16 vulnerable scenarios.

Доказательства

An LLM writing a convincing PoC string is not a finding.

Agents don't coordinate by vibes in a chat log. They claim explicit plan steps, attach real evidence, and update state through a revisioned store.

Egress fails closed. The declared scope is compiled into a netfilter allowlist inside the sandbox's own per-run network namespace.