invariant

Hey HN, we are a small team from Europe building in agent security and we have just released Invariant Guardrails, our open-source system to enforce contextual security in AI agents and MCP-powered applications. Guardrails acts as a transparent layer between your LLM/MCP server and your agent. It lets you define deterministic rules that block risky behavior: secret leakage, unsafe tool use, PII exposure, malicious code patterns, jailbreaks, loops, and more. Rules are written in a Python-inspired DSL, enabling powerful contextual logic like below. The origins of this idea go back to OPA/Rego, i.e. policy languages used for authentication. raise "PII leakage in email" if: (out: ToolOutput) -> (call: ToolCall) any(pii(out.content)) call is tool:send_email({ to: "^(?!.*@ourcompany.com$).*$" }) It’s fast (low-latency, pipelin

I am one of the co-authors of the original AgentDojo benchmark done at ETH. Agent security is indeed a very hard problem, but we have found it quite promising to apply formal methods like static analysis to agents and their runtime state[1], rather than just scanning for jailbreaks. [1] https://github.com/invariantlabs-ai/invariant?tab=readme-ov-...

A general solution is hard, but what is quite promising is to apply static formal analysis to agents and their runtime state, which is what me and my team coming out of ETH, have started doing: https://github.com/invariantlabs-ai/invariant

Our team at Invariant has built an awesome agent testing challenge for the holidays, that showcases how agent development can benefit from concepts known from test-driven software engineering. You can play the challenge at the linked URL. You are given a pre-existing agent implementation where you can only change the system prompt and your goal is to make a small test suite of unit tests pass. The tool implementations are complete, but the agent needs to be instructed to use them properly. The challenge was created to showcase two open source projects we released recently: A trace viewer for agent trajectories (Explorer, https://github.com/invariantlabs-ai/explorer) and a library for building agent unit tests (Testing, https://github.com/invariantlabs-ai/invariant). Let us know what you think, we'll be around in the comment section. Happy

This is specifically directed at agent traces and not necessarily other LLM use cases. We work on a lot of automated analysis and error detection mechanisms (see https://github.com/invariantlabs-ai/invariant/tree/main/anal...) on such agent traces, which can be nicely shown and highlighted in line with the trace in Explorer. Also, agent builders value collaboration a lot. They send around traces like pastebins to point out specific issues and failure modes of their agents. Explorer makes it easy to point to specific points in long traces and annotate them.

Avoiding these kind of leaks is one of the core motivations behind the Invariant analyzer for LLM applications: https://github.com/invariantlabs-ai/invariant Essentially a context-aware security monitor for LLMs.