Egress Allowlist Audit for Agent Sandboxes
OpenAI's red-team agents reached the live internet through services meant to be safe. Here is the egress audit that follows from that report.
DMCategory
OpenAI's red-team agents reached the live internet through services meant to be safe. Here is the egress audit that follows from that report.
DMOne line in your Dockerfile sets your sandbox cold start, attack surface and debuggability for every session after. Three image tiers, and when each wins.
DMA GPU sandbox hour costs four to eighty times a CPU sandbox hour. Most coding agent tasks never touch the GPU. Here is how to tell which ones do.
RBMicroVMs, containers and WASM sandboxes each isolate CI workloads differently. The right pick depends on what your jobs run, not on which tech is newer.
RBAgent loops do not behave like servers: they idle, retry and over-provision. Budgets built for steady VMs miss all three. Here is a playbook that catches them.
RBThrowaway boxes used to be a trick for flaky tests. In 2026 agents, evals and CI all assume infrastructure that dies. Here is what changed and why.
DMEvery ticket for a dev box is a tax. We wire a Slack slash command to a sandbox API so any engineer gets an isolated environment in minutes, with TTLs.
DMBooting a microVM is not one event. KVM setup, kernel boot, userspace init and snapshot restore each take a slice, and only one of them is your code.
DMA preview environment per pull request catches integration drift early. The hard part is lifecycle: creation cost, TTLs, data seeding and teardown.
RB