I build backends and guardrails for AI agents. Now I’m going a layer down: C++ and Linux.

A scanner tells you a vulnerable function exists somewhere in your dependencies. It doesn't tell you whether your own code can ever reach it. Published industry analysis puts roughly four in five 'critical' findings out of reach of the app that ships them, but that's an industry number, not a verdict about your repo. ReachGate walks the code graph to check whether a path actually exists. If it runs out of budget before it can be sure, it returns UNKNOWN rather than marking something safe it never proved. Every result exports as OpenVEX and SARIF with an sha256 manifest, so you can re-check it offline.

// two public merge requests in GitLab's own namespace, pipelines and both verdicts visible · open on gitlab

read the case →

This is the one I put on PyPI. It sits where an agent fetches a web page and checks whether the page was actually read, or whether the request just came back with a 200. A 200 only means bytes arrived, not that you got the page. If it can't confirm a real read, it stops the agent instead of letting it answer anyway. No model, no API keys. One test in the suite deliberately tries to slip a fake read past it, and fails.

// 45 tests, no network, no model · verify on pypi

read the case →

Before an agent does something costly, like issuing a refund, TrustGate looks at the data behind the decision: is it fresh, does it still match the contract it was written against, is the source even connected? An agent can be allowed to act and still be acting on bad data. It returns allow, approval-required, or block, with the evidence attached.

// recorded walkthrough, Cloud Run with BigQuery evidence checks · watch

read the case →