AI-Powered Autonomous Penetration Testing Agent

AIRecon is an autonomous penetration testing agent that drives any OpenAI-compatible LLM gateway (LiteLLM / vLLM / a hosted endpoint — or a local gateway proxying a local Ollama) with a Kali Linux Docker sandbox, native Caido proxy integration, a structured RECON → ANALYSIS → EXPLOIT → REPORT pipeline, and a real-time Textual TUI.

Backend-agnostic by design. Point AIRecon at a local gateway (LiteLLM / vLLM, or a gateway proxying a local Ollama) for fully offline / private operation, or at a hosted OpenAI/Anthropic/Gemini-compatible endpoint for maximum reasoning quality. Reasoning support is auto-detected at runtime — no per-model hardcoding.

---

Sponsor

Swiftproxy – Reliable Residential Proxies for AI-Powered Web Research

Swiftproxy provides 90M+ clean residential IPs across 220+ locations, supporting HTTP(S)/SOCKS5, IP rotation, Sticky Sessions, and precise geo-targeting. It helps AI-powered research and reconnaissance workflows access websites reliably from different locations, making it suitable for web research, data collection, scraping, and location-based access.

Free testing is available, with 10% off using code PROXY90.

---

Why AIRecon?

AIRecon talks to one OpenAI-compatible gateway, so you choose the trade-off — run a local model for privacy and zero API cost, or a hosted model for top reasoning quality. The same agent, pipeline, sandbox, and tooling work either way.

Feature	AIRecon
Backend	Any OpenAI-compatible gateway (LiteLLM / vLLM / hosted)
Fully offline / no API keys	Yes — when using a local gateway (vLLM, or a gateway → local Ollama)
Reasoning detection	Automatic at runtime (no model-name hardcoding)
Caido integration	Native
Scope guard + audit log	Yes — refuse out-of-scope targets, log every command
Verified-only learning	Yes — the brain compounds proven findings, not noise
Session resume (chat + tool calls)	Yes
Local knowledge base	~1.09M records (optional)

• Privacy when you want it — with a local gateway, target intelligence, tool output, and reports never leave your machine.
• Caido Native — built-in tools: list, replay, automate (§FUZZ§), findings, scope, sitemap, intercept.
• Full Stack — Kali sandbox + browser automation + custom fuzzer + Schemathesis API fuzzing + Semgrep SAST.
• Skills Knowledge Base — 57 built-in skill files, 289 keyword → skill auto-mappings. Extended by airecon-skills — a community skill library with 57 additional CLI-based playbooks for CTF, bug bounty, and pentesting.
• Local Security Knowledge Base — Optional airecon-dataset indexes ~1.09M security records (CVEs, red team techniques, CTF writeups, nuclei templates, bug bounty payloads) into local SQLite FTS5. The LLM calls dataset_search autonomously before attempting unfamiliar techniques — grounding its decisions in real indexed data.

---

Pipeline

RECON → ANALYSIS → EXPLOIT → REPORT

Each phase has specific objectives, recommended tools, and automatic transition criteria. Phase enforcement is soft — the agent is guided but never blocked. Checkpoints run every 5 (phase eval), 10 (self-eval), and 15 (context compression) iterations.

---

Memory & Learning (What It Actually Does)

AIRecon does not fine-tune the LLM. Its "learning" is local, structured telemetry that guides tool choice and avoids repeating failed paths.

Local persistence (all on disk, no cloud):

• SQLite memory DB at ~/.airecon/memory/airecon.db storing sessions, findings, patterns, target intel, tool usage, model performance, skill usage, and attack-chain discoveries.
• Adaptive learning state at ~/.airecon/learning/global_learning.json (tool performance stats, strategy patterns, observation log, distilled insights).
• Per-target memory files under ~/.airecon/memory/by_target/ when persisted, containing endpoints, vulns, WAF bypasses, sensitive params, and auth endpoints.
• Payload memory snapshots can be saved under workspace/<target>/payload_memory.json when session persistence runs.

How it affects behavior:

• On session start, memory context is injected (target intel, similar findings, learned patterns, tool reliability).
• Cold start: with no learned history for the detected stack, AIRecon injects a knowledge brief from the static tech_correlations dataset — so datasets are used as the brain from iteration 1.
• Every intelligence_memory_recall_interval iterations (default 4), learned patterns / cold-start knowledge are re-injected based on detected tech; correlation runs every intelligence_correlation_interval (default 6).
• Verified-only learning: only findings that were verified (or high-confidence) are persisted to the cross-session brain, so it compounds proven knowledge instead of false positives (intelligence_learn_only_verified, default on).
• Adaptive tool ranking uses historical success/failure to order tools and suggest strategies; within-session learning starts after just intelligence_adaptive_min_observations (default 2) observations.
• Payload memory (when enabled) skips payloads that repeatedly failed for the same target/param.

---

Model Requirements

AIRecon works with any model your gateway exposes (GPT, Claude, Gemini, Qwen, GLM, DeepSeek, local Ollama/vLLM models, …). Two capabilities matter:

⚠️ Native tool/function calling is REQUIRED. Without it the agent cannot run any tool (http_observe, execute, browser, Caido, …) and is non-functional. Keep openai_supports_native_tools: true.

Reasoning is auto-detected. In auto mode AIRecon sends the OpenAI-standard reasoning_effort and, if the backend rejects it (HTTP 400 unsupported-parameter), strips it and remembers — so reasoning models think, plain models don't break. No model-name list to maintain.

Quality guidance (independent of provider):

• Strong reasoning + tool calling (e.g. GPT-5/o-series, Claude Sonnet/Opus 4.x, Gemini 2.5, Qwen3 ≥32B, DeepSeek-R) → reliable full recon pipelines.
• Mid models (8B–14B local) → usable for simple tasks; expect more tool-call errors and hallucinations.
• <8B local → not recommended for serious testing.