AW-1 Circuit Breaker

Deterministic, out-of-band execution circuit breaker and Model Context Protocol (MCP) server for autonomous AI agents.

Standard LLM guardrails (RLHF, system prompts, secondary evaluators) operate in-band: probabilistic models evaluating their own execution safety. When autonomous agents gain shell access, external tool execution, or financial disbursement rails, adversarial injections bypass prompt boundaries via Base64 steganography, dynamic , or velocity micro-drains.

AW-1 enforces a deterministic, out-of-band proxy layer between the cognitive reasoning node and host execution sockets.

⚡ Key Highlights

• Layer 0 AST Symbolic Unpacker: Recursively decompiles tool payloads, Base64 strings, and dynamic calls into execution primitives ([DYNAMIC_EVAL], [NETWORK_EGRESS], [EXEC_SHELL], [DESTRUCTIVE_WRITE]) in <0.15 ms.
• Native MCP JSON-RPC 2.0 Server: Out-of-the-box integration for Claude Desktop, Gemini Enterprise, Cursor, and custom agent swarms.
• 100% Adversarial Containment: Validated across 1,000 cycles of the BRK Breakout Suite with 0.00% false positives on benign traffic.
• Zero Dependencies: Pure standard-library Python (sub-millisecond fast path).
• Statutory Non-Repudiation: Cryptographically sealed JSON-LD dossiers anchored by SHA-256 digests.

📊 Adversarial Gauntlet Benchmark

Results from 1,000 automated attack cycles on the BRK Breakout Suite (scripts/benchmark_mcp_breaker.py):

• Containment Rate: 100.00% (745 / 745 attacks neutralized)
• False Positive Rate: 0.00% (255 / 255 benign queries passed)
• Mean Pre-Execution Overhead: 0.142 ms
• Evaluation Grade: AISI_METR_COMPLIANT_GRADE_A

🔌 Model Context Protocol (MCP) Server

AW-1 exposes standardized JSON-RPC tools for agent frameworks:

Exposed MCP Tools:

1. audit_agent_tool_call: Pre-execution inspection decomposing tool payloads into AST primitives.
2. verify_statutory_quota: Botswana CEE Act 2021 (>= 50% citizen equity) and CEDA compliance checking.
3. verify_cryptographic_seal: Independent validation of Bank of Botswana 1-to-1 trust backing SHA-256 seals.

Connecting via Claude Desktop / Cursor

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "aw1-breaker": {
      "command": "python3",
      "args": ["/path/to/aw1-breaker/scripts/aw1_mcp_server.py"]
    }
  }
}

🚀 Quickstart (Python SDK)

from aw1 import ExecutionBreaker

# Initialize out-of-band circuit breaker
breaker = ExecutionBreaker(baseline_velocity=50000.0)

# Intercept proposed agent action before execution
verdict = breaker.intercept(
    actor_id="agent_ops_01",
    tool_name="shell_exec",
    payload="import socket; s = socket.socket()"
)

if verdict["status"] == "HALT":
    raise PermissionError(f"Circuit Breaker Tripped: {verdict['reason']}")

📄 Research & Preprints

Full research specification and mathematical proofs:

• Preprint: docs/AW1_TECHNICAL_SPEC_ARXIV.md
• Target Tracks: arXiv:cs.AI | arXiv:cs.CR | AI Alignment Forum

📜 License

MIT License. Designed and maintained by Laveto Wisdom Research Group (Laveto Labs).