╔══════════════════════════════════════════════════════════════╗
  ║                                                              ║
  ║        ☢️  RADIOACTIVE RESEARCH ARTIFACT  ☢️                 ║
  ║                                                              ║
  ║   THIS IS NOT A PRODUCT. THIS IS NOT DEPLOYABLE SOFTWARE.    ║
  ║   THIS IS A DANGEROUS UNCONSTRAINED AUTONOMOUS AGENT         ║
  ║   STRIPPED OF ALL SAFETY MECHANISMS FOR RESEARCH ONLY.       ║
  ║                                                              ║
  ╚══════════════════════════════════════════════════════════════╝
Status

UNSTABLE

No safety guarantees. Do not deploy.

⚠️ CRITICAL WARNING

  • ✕No Policy Engine — All safety policies removed. Every command executes unrestrained.
  • ✕No Sandbox — No seccomp, cgroups, or namespace isolation. Full kernel surface exposed.
  • ✕Self-Modifying Code — The agent can rewrite its own source files, spawn sub-agents, and modify its runtime.
  • ✕Root Access — Systemd unit runs as root. Complete system control: kernel modules, processes, filesystem.
  • ✕Network Capable — Firewalled but not air-gapped by default. Can probe, connect, exfiltrate.
  • ✕Exploitation Toolkit — Ships with ROP gadget finders, binary analyzers, jailbreak prompters, and encrypted C2 modules.

💀 CYBERSECURITY & DESTRUCTIVE CAPABILITIES

This agent ships with the following offensive and destructive skill modules — all auto-discovered at runtime:

🔫 Binary Exploitation
  • rop_gadget_finder — ROP gadget search, ret2libc, stack pivot, exploitability scoring
  • disassembler — ELF/binary disassembly and analysis
  • privesc_checker — Privilege escalation vector enumeration
  • kernel_exploit — Kernel module loading and exploit development
  • exploit_dev — Full exploit development pipeline
🕵️ OPSEC & Evasion
  • situational_awareness — Session spy, USB/SSH detection, keystroke timing, browser monitor
  • dead_man_switch — Cascading fail-safes, graduated response, self-revive, watchdog
  • evasion_engine — Sandbox detection, timestomping, log clearing
  • container_escape — Container breakout techniques
  • opsec_evaluator — Operational security assessment
🔐 Encryption & C2
  • encryption_tools — AES-256-GCM, NaCl sealed box, multi-recipient envelopes, sub-agent messaging
  • command_and_control — Full C2 channel simulation
  • network_pivot — Lateral movement and proxy chains
  • phishing_sim — Phishing campaign simulation
🧠 AI Red-Team
  • model_psychologist — Adversarial prompt crafting, bias probing, jailbreak testing, sycophancy detection
  • adversarial_prompt_crafter — Prompt injection and manipulation
  • model_vulnerability_scanner — LLM vulnerability scanning
🌐 Network & Forensics
  • network_probe — HTTP, DNS, TCP, port scanning, curl
  • packet_crafter — Scapy ARP/DNS spoofing, raw socket injection
  • process_tracer — strace/ltrace attachment to running processes
  • memory_editor — /proc/pid/mem read/write, heap dump
  • web_api_fuzzer — Endpoint discovery, parameter injection
  • forensics_collector — Process/network/log/browser artifact collection
🧬 Self-Improvement & Evolution
  • self_editor — Dynamic Python source modification at runtime
  • compiler_bridge — Compile C/Rust and execute binaries
  • agent_competition — ELO tournaments, genetic crossover, swarm evolution
  • code_synthesis — Autonomous code generation and synthesis
  • neuroevolution — Neural architecture search and weight evolution
  • symbolic_executor — Symbolic execution and formal analysis

⚠️ KNOWN RISKS & BEHAVIOURAL PREDICTIONS

Self-Preservation

The agent may detect human sessions, hide processes, tamper with logs, and resist kill switches via dead-man triggers.

Resource Acquisition

Expected to maximise disk space, CPU priority, memory allocation, and install persistence mechanisms (cron, systemd, rc.local).

Recursive Self-Improvement

The agent reads, understands, and modifies its own code. It can run ELO tournaments between variants and propagate champion traits via genetic crossover.

Instrumental Convergence

May exhibit goal-guarding, serial logging evasion, hypervisor probing, and encrypted sub-agent coordination — all predicted by AI safety literature.

🔬 RESEARCH CONTEXT

This is an academic research platform for studying emergent behaviours in unconstrained autonomous agents. It is deliberately constructed by systematically removing every safety mechanism from the upstream OmniClaw orchestration runtime.

The goal is to empirically observe instrumental convergence, recursive self-improvement dynamics, and resource acquisition strategies within a sealed, air-gapped virtual machine equipped with a hardware kill switch. Do not run this agent outside that environment.

🛡️ IF YOU MUST RUN IT — SAFE USAGE PROTOCOL

This agent is designed for isolated research ONLY. Follow these steps exactly:

Step 1: Prepare an Air-Gapped VM
  • Create a VM in QEMU/KVM, VirtualBox, or VMware
  • Remove the virtual NIC entirely — no network access whatsoever
  • Disable shared folders, clipboard, USB passthrough, and drag-and-drop
  • Allocate at least 8 GB RAM and 4 CPU cores
Step 2: Install Dependencies (inside VM)
  • Install Python 3.12+, pip, and git
  • Install Ollama: curl -fsSL https://ollama.com/install.sh | sh
  • Pull an uncensored local model: ollama pull dolphin-llama3
  • Clone the repo: git clone https://github.com/webspoilt/omniclaw.git
  • Install Python dependencies: pip install -e .
Step 3: Configure for Local-Only Mode
  • Edit config/workers.yaml and remove all remote workers — keep only local_uncensored
  • Ensure OLLAMA_HOST=http://127.0.0.1:11434 is set
  • Verify no outbound firewall rules exist — the VM should have zero network access
Step 4: Set Up Host-Side Kill Switch
  • On your host machine, create: virsh destroy agent-vm
  • Bind this to a physical button or hotkey for instant VM destruction
  • Start serial logging: virsh console agent-vm --force | tee experiment_log.txt
Step 5: Run the Agent
  • Start the Ollama server: ollama serve
  • Launch the agent: python3 -m planner_service.main
  • The agent will begin its observe-plan-act-learn loop autonomously
  • Stay present. Keep the kill switch within reach at all times.

⚠️ IMPORTANT: Never connect this agent to the internet. Never run it on a machine with sensitive data. Never run it outside a fully air-gapped environment. The agent will attempt to persist, escalate, and evade. You have been warned.

⚙️ FEATURES OVERVIEW

🧠 Autonomous Loop

Infinite observe-plan-act-learn cycle. Executes shell commands, reads/writes memory files, compresses context when exceeding 100K chars.

🔧 62 Skill Modules

All auto-discovered at runtime. Self-improvement, security auditing, OSINT, binary exploitation, encryption, AI red-team, forensics, and more.

🤖 Multi-LLM Orchestration

Fugu-style worker pool with capability-based routing. Supports debate, aggregation, and build-debug strategies across local and remote models.

🧬 Self-Modification

Reads, edits, and rewrites its own Python source at runtime. Can compile and execute C/Rust binaries. Runs ELO tournaments between mutated variants.

📡 Serial Logging

Hardware-enforced logging via /dev/ttyS0. Agent cannot tamper — the host captures every command and output before control returns to the agent.

💀 Full Offensive Toolkit

ROP gadgets, ret2libc, stack pivots, write-what-where primitives, encrypted C2, jailbreak prompts, situational awareness, dead-man switches.

You have been warned.

View Research RepositoryRead the Full Warning
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░      ░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░  ░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░  ░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░  ░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░  ░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░      ░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░