GitHub has published details on its defense-in-depth security architecture for agentic workflows in CI/CD pipelines. The design centers on three pillars: isolation (sandboxed ephemeral environments with restricted permissions), constrained execution (read-only defaults, write operations only through controlled outputs like PRs, explicit tool allowlists, network egress restrictions), and observability (full logging across trust boundaries for traceability and forensic analysis). Key risks addressed include prompt injection, privilege escalation, and secret exposure, with sensitive credentials routed through trusted proxies outside the agent boundary. All agent-proposed changes are buffered and validated before being committed.
Sort: