chu0119/Pojun
AI 自主渗透测试 · 红队作战平台 | 多 Agent 自主推进 · 漏洞情报运营 · 后渗透与 C2 | LLM 三通道轮询 | 全程证据留痕 | 原百度攻防赛冠军项目独立演进版
What it solves
Pojun is an AI-driven autonomous penetration testing platform. You give it a one-sentence target (a domain, IP, or company), and it automatically runs the full security assessment pipeline: target decomposition → reconnaissance → vulnerability discovery → exploitation → post-exploitation → report generation. Humans can step in and correct course at any point.
It is not a wrapper around a scanner. The core design is to make the AI work like a disciplined penetration testing engineer. Two hard gates keep results honest:
- Goal gate — every goal must carry a verifiable acceptance criterion (e.g. "an unauthenticated endpoint returns 200 with business data"). A task can only be marked complete when all goals are proven met; otherwise it is honestly marked "incomplete" in the report.
- Evidence gate — critical/high vulnerabilities must come with mechanically reproducible evidence (HTTP status codes, command output, traffic comparisons). Pure version-matching "might exist" findings never make it into the deliverable report.
It also keeps full traceability: all HTTP/S traffic passes through a MITM proxy and is recorded request-by-request, so every conclusion can be traced back to its evidence.
How it works
Pojun is a Go monolith backend with an embedded Next.js frontend and PostgreSQL storage. The engine runs multiple AI agents:
- planner — the only intent generator. It holds a shared to-do list across wake-ups and dispatches 0–4 intents per round, serializing exploitation chains (injection point → credentials → lateral movement) by dependency.
- worker ×N — agents that claim intents from a queue, execute real tools (all traffic logged through the MITM proxy), and write results back to the graphs, which wakes the planner for the next round.
- mainagent — handles conversations, report writing, and immediate tool-triggered tasks.
State is split into two graphs stitched together by anchors:
- Asset graph (global, shared across tasks) — the truth store for root domains, subdomains, IPs, services, apps, and endpoints, with company ownership and authorization boundaries attached.
- Exploration graph (per task) — the progression chain from goal → intent → fact/finding, with five edge types (spawns, derived_from, yields, proves, covers) forming lineage so every finding can be traced back to its originating clue.
Key mechanics:
- Attack chain continuation — confirming a vulnerability is not the end. The platform auto-dispatches next-hop intents by vulnerability class: SQL injection → data exfiltration → login state → backend lateral movement; file upload → webshell capability confirmation; unauthorized API → privilege escalation comparison → data impact assessment.
- Deduplication — findings carry strong keys (cve/root/endpoint/param) plus weak keys (asset_class/name); new reports auto-merge into existing findings with a merge timeline, and cross-task duplicates are suppressed by default.
- Write-back guard — recording facts or findings anchored to assets outside the task's declared scope is rejected at the data layer.
- Three-layer approval gates —
deny(hard-blocked, no model call),judge(model decides based on context, strict JSON parsing),ask(human approval for platform capability changes like opening listener ports). - Vulnerability intelligence — continuously syncs NVD (420k+ CVEs), CISA KEV in-the-wild exploitation, GitHub Advisory, and PoC repositories. High-severity (CVSS ≥ 9.0 or KEV) items are pulled every 30 minutes and auto-matched against asset fingerprints, with hints attached to tasks.
- Red team capabilities — a reverse-shell/OOB callback receiver, C2 egress in three modes (direct / SSH relay / manual), and a six-phase post-exploitation methodology.
- LLM routing — multiple model channels with polling, circuit breaking, failover, and role-based model assignment (strong reasoning for planner, mid-tier for workers, cheap for judge).
Who it's for
Security professionals running authorized penetration tests, red team operations, and SRC (security response center) hunting. It is deployed on Linux (Debian/Ubuntu/RHEL/Kali), macOS 13+, or Windows via Docker Desktop. The full toolchain (66 tools) is bundled in the Docker image; on native installs, missing tools are honestly shown on the "host environment" page and the agent works around them. A single-skill version (pojun-skill) is also available for any AI agent.
Highlights
- One-command install — a curl pipe script handles dependencies, PostgreSQL, service registration, and health checks on Linux and macOS; Docker Compose works on all platforms with the toolchain fully inside the image.
- Honest completion — tasks only reach "done" when all goals with verifiable criteria are met; otherwise they are marked "timeout/incomplete" in the list, detail view, and report.
- Evidence-backed findings — critical/high vulnerabilities require mechanically reproducible proof, with four classes of bound traffic evidence (baseline/proof/verification/supporting), sha256 versioned snapshots, and adversarial self-checks (reproduce with different identity/parameter order/client).
- Professional deliverables — Word reports with cover page, table of contents, severity coloring, retest records, traffic evidence, and remediation roadmap (hand-written OOXML, zero dependencies), plus Markdown and per-vulnerability docx zips.
- Real-world baseline — 12 authorized tasks, 56,000+ tool calls, 295 confirmed vulnerabilities (28 critical / 140 high), 2,752 asset nodes, fully unattended.
- 22 built-in skills, 10 MCP presets, 66 host tools — covering web/API testing, cloud, Windows AD, source code audit, LLM security, mobile reverse, red team post-exploitation, and SRC hunting knowledge.
- Cost control — multi-channel LLM polling with circuit breaking and failover; measured 2,725M tokens with 93% cache hit rate; free rate-limited channels can still finish tasks.
Related
- Project
- Project
- Project
- Project