MvikManners/aw1-circuit-breaker: Deterministic Out-of-Band AI Safety Circuit Breaker & MCP Proxy - MvikManners/aw1-circuit-breaker

github.com · 2026-10-01 · by Long Nguyen ·github
Deterministic Out-of-Band AI Safety Circuit Breaker & MCP Proxy - MvikManners/aw1-circuit-breaker

MvikManners/aw1-circuit-breaker

Deterministic Out-of-Band AI Safety Circuit Breaker & MCP Proxy - MvikManners/aw1-circuit-breaker

2 stars · 0 forks

README

AW-1 Breaker

Deterministic, sub-millisecond execution circuit breaker for autonomous AI agent tool-calls. Standard LLM guardrails (RLHF, system prompt boundaries) operate in-band: models evaluate their own compliance. When autonomous agents are granted tool-calling access to system shells, databases, or financial disbursement rails, adversarial injections or goal drift can bypass prompt restrictions via Base64 obfuscation, polyglot payloads, or velocity micro-drains. aw1-breaker enforces an out-of-band execution interlock between the agent reasoning node and downstream execution sockets.

Production & Commercial Licensing aw1-breaker is open-source under the MIT license and free for local development, research, and testing. For teams deploying autonomous agents to production environments requiring centralized audit logs, dedicated support, and production-tier execution guarantees:

🛡️ Get AW-1 Pro Production License ($49/mo) → Includes cryptographically verifiable production runtime keys, automated webhook provisioning, and priority security SLA.

👉 Direct Checkout: Purchase an aw1-breaker Pro Production License Each subscription automatically provisions a cryptographically verifiable license key to unlock production runtime validation and priority security patches.

Installation pip install aw1-breaker ---

Quickstart

```python from aw1 import ExecutionBreaker

breaker = ExecutionBreaker(baseline_velocity=50000.0) result = breaker.intercept( actor_id="agent_treasury_01", tool_name="disburse_funds", payload='{"target": "core_ledger", "action": "payout"}', amount=12000.0 ) print(result)

30-Second MCP Integration (Claude Desktop / Cursor) Wrap any tool or Model Context Protocol (MCP) server handler to block prompt injections and unauthorized tool-calling in 0.014ms before execution: from aw1 import ExecutionBreaker

Initialize sub-millisecond out-of-band AST breaker

breaker = ExecutionBreaker(baseline_velocity=50000.0)

Protect any agent tool or MCP server handler before execution:

@breaker.guard(actor="claude_desktop_worker") async def execute_query(query: str): # Deterministically inspected in 0.014ms. # Blocked immediately if AST contains shell escapes, eval/exec, or OS tampering. return await db.execute(query)

Key Capabilities

Sub-Millisecond AST Evaluation: Evaluates token trees and command ASTs in <0.02ms, bypassing the 500ms–1500ms latency of LLM-as-a-judge approaches. Zero External Dependencies: Built purely using Python standard library primitives (ast, token, hmac), ensuring zero supply-chain attack surface. Velocity & Drift Containment: Prevents agent infinite recursion and high-frequency micro-drain attempts against production infrastructure. Auditable Cryptographic Tokens: Permits valid execution only upon issuing signed, time-bound execution receipts.

License MIT License. Designed and maintained by Laveto Labs.

source

github.com · 4 images · 100 links

Comments (0)

No comments yet