Egress control decides which hosts an agent's commands can reach, and it is the main defense against exfiltration after a successful injection. Common patterns are a filtering proxy with a domain allowlist, firewall rules inside a container, and vendor-managed allowlists for cloud agents. Watch for side channels such as DNS lookups and allowlisted domains that host user content.
Read in this order
- OfficialClaude Code: Configure the sandboxed Bash tool (network)Start here: how a domain allowlist is enforced through a proxy.
- ArticleClaude Code: Data exfiltration with DNS, CVE-2025-55284 (Embrace The Red)A real exfiltration through DNS that a naive allowlist missed.
- VideoBreaking the Lethal Trifecta (Without Ruining Your Agents) ([un]prompted 2026)Relates egress restrictions and confirmation flows to production controls. 19 min.