Anthropic publishes security guide for locking down Claude Managed Agents sandboxes
brada · x · 2026-10-14
Anthropic's brada notes that hosting sandboxes for highly capable, persistent models is "not a weekend project," and the company published a guide to securing Claude Managed Agents deployments.
- Credentials stay in a vault; the sandbox only sees placeholders, injected as request headers with limited networking
- Network access is limited to required hosts; websearch/webfetch get per-tool domain allowlists
- Permission policies: auto-approve read-only tools, alwaysask for irreversible actions
- Smallest tool set by default; agents reading untrusted input must not hold always-allowed write/execute tools
- Per-session budget caps and optional offline transcript review
A directly actionable hardening checklist for anyone building agents.
Related event: Anthropic Releases Security Hardening Guide for Claude Managed Agents(2 posts)→
More from coding & agent
- Training repo-level code editing into weights to ditch heavyweight agent scaffolds — burkov · 2026-10-14
- UCLA's PANINI-DR hits 73.3% on BrowseComp-Plus, beating best deep-research harness by 16 points — M. Yigit Turali · 2026-10-14
- Mattermark creator launches StartupGraph: startup data API and MCP server for AI agents — Scobleizer · 2026-10-14
- 10 open-source repos to test your AI agent before your users do — victor_explore · 2026-10-14
- With superhuman coding agents, changing habits is still the hardest engineering problem — daniel_mac8 · 2026-10-14
- Robo Use launches: one API for agents across 32 robot embodiments — joemeno · 2026-10-14