Skip to content
Computing Life · Share · Yage

Agent Security Has No Universal Sandbox: A Five-Layer Interception Chain from Intent to Outcome

Using a case where npm test hides a malicious subprocess that steals SSH keys, this piece breaks Agent security into five layers: internal activation probing, chain-of-thought auditing, structured tool-call authorization, static command inspection, and kernel-level sandboxing plus resource-side immutable boundaries. The core insight: layers closer to the model understand intent better but are easier to bypass; layers closer to the OS enforce hard limits but understand zero business semantics. The post explicitly states that J-lens mind-reading is only probabilistic early warning, chain-of-thought can be unfaithful, the MCP gateway can't see dynamic subprocesses spawned by npm test, and static rules miss runtime child processes—only Landlock/seccomp or gVisor/Firecracker isolation finally blocked the exfiltration. It also debunks three sandbox myths: cutting the network doesn't make you safe, VMs can leak cloud credentials via metadata services, and detect-and-kill loses the race to exfiltration.

Why it matters: A malicious subprocess hidden in npm test stealing SSH keys shows why agent defense needs five layers, from activation probing to kernel sandboxes.

Read the original ↗Export Markdown