P5 — ATLAS (Autonomous Pentest Pipeline)

What It Does

Point ATLAS at an HTB/CTF machine IP. It runs a six-stage pentest pipeline autonomously — scans ports, enumerates every service, attacks the web layer, exploits, escalates to root, and generates a Markdown writeup. Human-in-the-loop checkpoints let you review and approve every action, or run fully autonomous with --auto.

Pipeline

Target IP
      ↓
  Recon         nmap fast → deep, NSE vuln scripts, parallel whatweb
      ↓
  Enumerate     SMB + FTP + LDAP + SNMP in parallel, hydra brute on found users
      ↓
  Web           nikto + gobuster simultaneously per target, vhost ffuf
      ↓
  Exploit       searchsploit + LLM plan → execute → shell + flag detection
      ↓
  PrivEsc       SSH exec: SUID, sudo, caps, cron, kernel, LinPEAS → root flag
      ↓
  Report        Full Markdown writeup, auto-commit to ctf-lab

Elite Features

• Tiered LLM — Haiku for cheap recon/enum analysis, Sonnet for exploit/privesc planning. Reduces per-run API cost 60–80% versus Sonnet-only.
• Prompt caching — Claude system prompts cached across all six agents (~90% cost cut on repeated system context).
• Retry + backoff — exponential backoff on rate limits and 5xx errors. Cache hit-rate telemetry printed at session end.
• NSE vuln scripts — free CVE detection: EternalBlue (MS17-010), Shellshock, Heartbleed, MS08-067, smb-vuln-*.
• Hydra brute-force — auto-fires against SSH/FTP when usernames are enumerated. Top-500 rockyou passwords, -f stop-on-hit.
• SSH-based PrivEsc — sshpass runs 18 Linux enum commands (SUID, sudo -l, capabilities, cron, kernel, docker group, bash history). Optional LinPEAS via SSH pipe.
• HTB flag auto-capture — recognizes HTB{...} and 32-hex legacy flags. Reads /root/root.txt on uid=0 detection.
• Session portfolio — atlas.py --list-sessions shows every past run with root/user checkmarks.

Parallel Tool Execution

Every stage runs independent tools simultaneously via concurrent.futures.ThreadPoolExecutor:

Typical time saving: 60–70% reduction in tool execution.

Autonomy Dial

--auto-risk controls how much the agent handles itself before pausing:

Even in auto mode, every checkpoint still renders FOUND / PLAN / WHY / LOOK-FOR / COMMAND — nothing is a black box.

Human-in-the-Loop Checkpoints

Before every action ATLAS shows:

• FOUND — what was discovered
• PLAN — what will happen next
• WHY — attack reasoning
• LOOK FOR — what indicates success
• COMMAND — exact command that will run
• RISK — low/medium/high/critical

Choose: a=approve, s=skip, m=modify command, q=quit.

LLM Providers

Auto-tier map: recon/enumeration → Haiku 4.5, web/exploit/privesc/report → Sonnet 4.6.

Engineering Details

Prompt caching architecture — every agent extends BaseAgent with a shared SYSTEM_PROMPT. ClaudeProvider marks the system block with cache_control: ephemeral. Subsequent agent calls read from cache at ~10% the input price.

Thread-safe parallel output — parallel tool runs suppress per-line streaming (interleaved output is unreadable). Each tool writes its full output to a log file; results are collected after as_completed(). A print lock serialises completion notices.

Checkpoint-then-parallelize pattern — EnumerationAgent shows per-service checkpoints sequentially (preserving the interactive approval flow), collects approved tasks into a list, then fires them in a single run_parallel() call. No checkpoint is skipped; parallelism only applies to approved execution.

Session resume — full state (ports, credentials, findings, AgentResult list, LLM metadata) is persisted after each stage. atlas.py --resume --stage privesc picks up exactly where it left off with all context intact.

Shell detection fix — subprocess captures non-interactive output. The original regex looked for $ and # prompts that never appear post-hoc. Rewritten to match concrete RCE evidence: uid=N(user), /etc/passwd content, uname -a output, Windows version banner, meterpreter tags.

Usage

# Interactive (default) — pause at every checkpoint
atlas.py 10.10.11.100

# Autonomous below high risk
atlas.py 10.10.11.100 --auto --auto-risk high

# Local claw-core instead of Claude
atlas.py 10.10.11.100 --provider ollama --model hermes3:70b

# Resume from privesc stage
atlas.py 10.10.11.100 --resume --stage privesc

# Portfolio view — every past run
atlas.py --list-sessions

Output

Each run produces ~/atlas-sessions/<ip>/:

• *.txt — raw tool logs per command
• session.json — full state: ports, creds, findings, agent results
• writeup/<date>-<machine>.md — HTB writeup, auto-committed to ctf-lab
• atlas_privesc.sh — enum script if no SSH creds (fallback path)

Stack

Python 3.11+ · Anthropic SDK (caching enabled) · Ollama · rich · nmap · gobuster · nikto · ffuf · enum4linux-ng · smbclient · ldap-utils · snmp · hydra · sshpass · searchsploit