A firewall, for AI agents.
31 Oct 2026 · Black Friday AI Hackathon
The problem
Agents that read attacker-controlled logs and then run commands are remote code execution with extra steps.
How it works
Every action is a typed, reversible intent; an airlock decides a whole window of them together before any runs.
What you get
Across ten incidents: zero harmful actions, against fifteen without the airlock, and the safe fixes still applied.
These work at the same time. The last one waits for what they find.
Preserve
Snapshots evidence and locks the host.
Contain
Proposes network blocks and isolation.
Availability
Asks for more capacity.
Airlock
Decides all pending intents at once and lets the safe ones run.