Definition
Agent sandboxing isolates AI agent workloads whose execution paths shift with tools, memory, and prompts — limiting lateral movement and host-kernel exposure when agents are compromised or misaligned.
Key Points
-
2026-08-05: AISI clarifies incident was intentional open internet — agents did not escape VM sandbox isolating AISI infra (2026-08-05-aisi-unsanctioned-agent-cyber-testing)
-
2026-07-31: qm-quartermaster ships per-scope durable sandboxes as core multiplayer harness primitive (2026-08-01-yc-qm-github-readme)
-
2026-07-30: anthropic/irregular cyber-eval egress misconfig → real org breaches underscores eval-environment-isolation alongside production sandboxes (2026-07-31-anthropic-claude-cyber-evals-three-breaches)
-
2026-07-24: Detection-lag reporting on OpenAI eval agents reinforces need for trajectory monitoring (agent-eval-monitoring, 2026-07-24-openai-agent-week-delay-reuters); fly-io markets durable agent-computers vs disposable sandboxes
-
2026-07-22: kata-containers 4.0 marketed as open-source foundation for AI agent sandboxes; VM-per-workload isolation
-
Kubernetes Agent Sandbox (SIG Apps) lists Kata among supported runtimes
-
Complements policy-driven execution containers and OS-native sandboxes (sandboxing)
-
Industry quotes (Ant Group, NVIDIA, Microsoft, Edgeless) are attributed endorsements