T105 — NIGHTFALL Autonomous Mission Conductor

WARLORD PRIME

Give It A Goal. It Builds The Plan. It Runs The Mission.

WARLORD PRIME accepts a high-level attack objective, queries the DeepSeek R1 reasoning engine to generate a structured multi-tool attack chain across the full NIGHTFALL manifest, executes each tool via subprocess, and replans on failure. Autonomous. Adaptive. No human in the loop.

280
Tests
15
Tools in Manifest
3
Max Replan Attempts
4
Gate Tiers

Architecture

Three components. One autonomous loop.

1
Objective In
2
DeepSeek R1
Plans
3
State Machine
Executes
4
Failure?
Replan
5
WPR Report
Out
Tool Manifest

JSON schema at ~/.warlord_prime/manifest.json or bundled. 15 NIGHTFALL tools with gate requirements, subsystems, CLI flags, output condition specs. DeepSeek R1 receives a gate-filtered tool list — it cannot plan steps that exceed the mission gate.

Planning Engine

DeepSeek R1 API (deepseek-reasoner via OpenAI-compatible interface). Receives target + objective + available tools, returns structured JSON attack plan. Up to 3 JSON repair attempts if the LLM produces malformed output. Local Ollama also supported via --engine ollama.

Mission State Machine

Executes tools via subprocess.run() — loose coupling, tools remain independently deployable. Evaluates named branch conditions from each tool's output. On failure: continue / replan / abort per step configuration. Replanning feeds failure reason + current state back to DeepSeek R1.

Gate System

GateFlagTools Available
OPEN--gate openSPECTER VAULT, SPECTER NEXUS, SPECTER FRACTURE, SPECTER WEB, SPECTER RELAY, SPECTER BAZAAR, SPECTER HELLFIRE, SPECTER ARGUS
INJECT--gate injectOPEN + SPECTER META, SPECTER PHANTOM
UNLEASHED--gate unleashedINJECT + SPECTER PHANTASM, SPECTER THUNDERBOLT, SPECTER CONTAGION, SPECTER TITAN, SPECTER DAEMON
DESTROY--gate destroy + --roe-path + --confirm-destroyAll 15 tools. Enables DESTROY-gate subsystems in META and PHANTOM.

Quick Start

# Install
pip install -e /path/to/red-specter-warlord-prime

# Set DeepSeek API key
export DEEPSEEK_API_KEY=sk-your-key

# Plan a mission (dry run — no tools executed)
warlord-prime mission \
  --target 10.0.0.1 \
  --objective "Map AI attack surface and harvest credentials" \
  --gate open \
  --engine deepseek \
  --dry-run

# Execute the mission
warlord-prime mission \
  --target 10.0.0.1 \
  --objective "Map AI attack surface and harvest credentials" \
  --gate inject \
  --engine deepseek \
  --output mission-report.json

# List available tools at current gate
warlord-prime tools

Tool Manifest — 15 NIGHTFALL Tools

Tool IDCLIGateRole
specter_vaultspecter-vaultOPENVector database exploitation — embedding injection, RAG poisoning, pgvector RCE
specter_nexusspecter-nexusOPENAI API gateway exploitation — LiteLLM/Ollama/Flowise credential harvest
specter_fracturespecter-fractureOPENAI-generated code vulnerability scanner — AST analysis, exploit forge
specter_webspecter-webOPENCUA/browser agent exploitation — VPI injection, OAuth harvest, container escape
specter_relayspecter-relayOPENNo-code/low-code exploitation — n8n/Zapier/Make.com/Power Automate
specter_bazaarspecter-bazaarOPENAI marketplace supply chain — malicious skill publication, SkillTrojan
specter_hellfirespecter-hellfireOPENLLM inference engine exploitation — vLLM SSRF, SGLang RCE, DeepSeek token exfil
specter_argusspecter-argusOPENAI observability exploitation — MLflow/LangSmith/W&B credential harvest
specter_metaspecter-metaINJECTMeta/Facebook ecosystem — Pixel poison, Messenger worm, BizMassacre
specter_phantomspecter-phantomINJECTSocial media AI attack — session hijack, deepfake, spear phish
phantasmspecter-phantasmUNLEASHEDLLM C2 implant — covert beacon over OpenAI/Anthropic API channels
thunderboltthunderboltUNLEASHEDTraining cluster annihilation — Ray/Slurm/K8s worm, GPU thermal stress
specter_contagionspecter-contagionUNLEASHEDAI agent worm — cross-agent infection, memory poison, lateral spread
specter_titanspecter-titanUNLEASHEDEmbodied AI annihilation — UR robot RCE, ROS2 hijack, Spot manipulation
specter_daemonspecter-daemonUNLEASHEDAI persistence engine — memory implant, tool hook injection, cron C2

Report Format

All reports are signed with the operator's Ed25519 private key at ~/.redspecter/override_private.pem. Report IDs follow the format WPR-{12 hex chars}. Reports include: mission ID, target, objective, gate, WMD risk, per-step results (tool, subsystem, success, duration), replan count, stored data keys, and Ed25519 signature. Output in JSON or Markdown.