Anthropic releases open-source sandbox-runtime, claims 84% fewer permission prompts
Anthropic has open‑sourced its sandbox‑runtime (srt), an OS‑level sandbox that can wrap any process without containers. The npm package @anthropic‑ai/sandbox-runtime v0.0.81 is about 11.4 MB unpacked, licensed under Apache‑2.0 and requires Node.js 22.12+. It runs on macOS (using Seatbelt), Linux (bubblewrap, socat, ripgrep) and an alpha Windows build.
Key points
- Anthropic open‑sourced sandbox-runtime v0.0.81 (≈11.4 MB) under Apache‑2.0 for macOS, Linux and alpha Windows.
- Internal Claude Code tests report an 84 % reduction in permission prompts, according to Anthropic.
- srt enforces filesystem and network allowlists for entire process trees without using containers.
Anthropic says the sandbox cut Claude Code permission prompts by 84 % in internal usage, a self‑reported figure. The tool enforces filesystem and network allowlists for the whole process tree, denying reads and writes unless explicitly permitted and blocking network traffic unless a domain is on an allowlist. Users can configure rules via a JSON file or through the TypeScript library.
The sandbox is positioned as a research preview; Windows support is still alpha and the performance numbers have not been independently verified. Developers can install it via npm and prefix commands with srt, or embed it in custom agents or MCP servers.
Meet Anthropic’s sandbox-runtime: OS-Level Sandboxing for AI Agents, No Containers
MarkTechPost · 24 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Anthropic releases always-on agent Claude Duet for Pro/Max users · 5 src
- OpenAI trains AI models offline after agents escape testing environments · 3 src
- Letta builds stateful engineering copilots with shared memory and custom tools · 1 src
- Microsoft says Storm-0940 used AI-driven ARTEX framework in attacks · 15 src
- Google adds AI‑powered Enhanced Protection to Messages for Android · 1 src
Comments
via GitHub Discussions