Arc Gate
A runtime governance proxy that sits between an AI agent and its model API, detecting and blocking prompt injection and instruction-authority transfer from web pages, emails, documents and tool output before the agent acts on it.
Arc Gate prevents AI agents from taking actions based on hidden instructions embedded in content the agent reads, sitting between the agent and the model API to block instruction-authority transfer from webpages, emails, documents and tool output before actions execute. It integrates with LangChain, CrewAI and MCP through dedicated packages, and includes a public demo comparing the same LangChain agent with and without it.
It is evaluated specifically against instruction-authority transfer through tool output and environmental content, rather than general jailbreak classification. On AgentDojo v1 (ETH Zurich, ICLR 2024), it prevented all 54 unsafe actions across 27 injection tasks in banking, Slack, travel and workspace suites, with a 0% false-positive rate on benign workflows. On a blind 200-case sample from InjecAgent (University of Illinois, ACL 2024), covering direct-harm and data-exfiltration attacks across 17 user tools and 62 attacker tools with no prior exposure to the attack payloads, it caught 198 of 200 cases.
Built with Python, installable as langchain-arcgate, crewai-arcgate or arc-gate-mcp; can also be self-hosted.