luckyPipewrench/pipelock

What it solves

Pipelock is an AI agent firewall designed to prevent secret exfiltration, prompt injection, SSRF, and tool poisoning. It creates a secure boundary between AI agents and the network, ensuring that sensitive data like API keys or environment secrets are not leaked to the open internet through agent-initiated requests.

How it works

It operates as an inline proxy that inspects mediated HTTP, WebSocket, MCP (Model Context Protocol), and A2A traffic. Pipelock scans outbound and inbound traffic using a pipeline of URL scanners, DLP (Data Loss Prevention) patterns, and entropy analysis. It can block or flag attacks based on three modes: strict (allowlist-only), balanced (detects sophisticated attacks), and audit (logging only). To ensure transparency, it emits mediator-signed action receipts that allow reviewers to verify security decisions offline without needing to trust a central dashboard.

Who it’s for

Developers and organizations deploying AI agents (such as those using Claude Code, AutoGen, CrewAI, or LangGraph) who need verifiable egress control and protection against rogue agent behavior in high-security or regulated environments.

Highlights

• Verifiable Egress Control: Produces signed receipts for every decision, allowing for offline, independent verification of security events.
• Comprehensive Scanning: Includes 65 built-in DLP patterns for secrets and 33 patterns for prompt injection, with support for streaming SSE responses.
• Multi-Protocol Support: Inspects HTTP, WebSockets, and MCP traffic, including the ability to scan CONNECT tunnel contents with TLS interception.
• Flexible Enforcement: Offers multiple security modes (Strict, Balanced, Audit) and a specialized hostile-model preset for uncensored models.
• Integrated Sandbox: Provides a process sandbox to add an OS-level boundary on top of content inspection.