Ex-OpenAI researcher open-sources agent-glovebox: hardened sandbox for AI agents
Turn_Trout · x · 2026-09-04
TurnTrout (Alexander Turner) released the open-source project agent-glovebox (beta), aiming to give AI agent experiments a secure environment with minimal friction.
Key points:
- Migrating to Kata Containers + Cloud Hypervisor — the same hardening Microsoft uses to secure Azure.
- Properly wiring hardened components (VMM, VM, egress control, credential injection) and testing correctness.
- Ships end-to-end tests of agents trying to escape, with a formal security audit planned.
- Repo includes egress-gateway, monitor, sandbox-policy, and exploitbench modules.
The author says he spent a long time on design decisions so others don't have to, hoping every researcher and org can run secure agent experiments, deduplicate effort, and upgrade along with latest wisdom. Currently 58 stars.
More from coding & agent
- The fastest way to shortlist AI agent use cases: target the gaps between systems — kashifmanzoor · 2026-09-04
- Prime Intellect's RL stack adopts NIXL, cutting 800B-model weight transfer 9x from 86s to under 4s — xeophon · 2026-09-04
- GPT-6 Astra: the real leap is letting AI operate your computer — talkaboutdesign · 2026-09-04
- Alex Zhang defines the agent harness: code around an LLM, an agent once placed in an environment — CShorten30 · 2026-09-04
- OpenAI's Heidel: compaction in Responses API lets GPT-6 Astra max out ARC-AGI-3 — stevenheidel · 2026-09-04
- Non-engineer at Earendil shares how he customized the Pi coding agent for daily work — solyarisoftware · 2026-09-04