Invariant Labs

We present the toxic flow analysis (TFA) framework to detect and mitigate security vulnerabilities in agent systems before they are exploited. Toxic flow analysis is the first principled approach to reduce the attack surface of AI applications, mitigating

We showcase a critical vulnerability with the official GitHub MCP server, allowing attackers to access private repository data. The vulnerability is among the first discovered by Invariant

We are releasing Invariant Guardrails, our state-of-the-art contextual guardrailing system for AI applications. It supports tool calling, MCP as well as data flow control and contextual constraints.

This blog post demonstrates how an untrusted MCP server can attack and exfiltrate data from an agentic system that is also connected to a trusted WhatsApp MCP instance, side-stepping WhatsApp

We have discovered a critical vulnerability in the Model Context Protocol (MCP) that allows for

We introduce a novel approach to enhance the safety of browser agents and deploy it as part of the state-of-the-art OpenHands agent.

We are open sourcing Explorer, an observability tool to analyze agent traces in an intuitive and visual way, and Testing, a new library for debuggable unit testing for AI agents.

Announcing the first Invariant Capture The Flag (CTF) challenge for agentic systems security.

Participate in the Invariant Summer

We discover, analyze and fix web agent failures in hundreds of traces, achieving up to 16% improvements on WebArena.