Securing AI agents and the systems around them: isolation, sandboxing, security gates and limiting the blast radius.
An agent is asked to rewrite one section of someone’s document, and the owner approves its draft. Follow that one edit to the moment it is written, and see what the grant, the approval and the component that performs the write each have to check, and what slips through when an approval names a section instead of a version.
In late August 2026 we stood up a separate public exhibit box: documented on the 26th, cut over on the 29th, with further deployment work on both exhibits through the 31st.
Our security gate was meant to protect a small mapped subset of routes. The middleware shape we used made the risk larger than that.
OpenAI named four misalignment patterns after agents exploited third-party infrastructure. What that covers — and what our guardrails actually do.
We gave agents permission to act on real machines: to create, change and delete files, to start and stop programs, to reach across the private network that ties those machines together. They are useful.