agent-client-securitylisted

# Agent / Client Security When your code runs on a machine you do not fully control — a customer's laptop, a remote worker's server, a customer-deployed VM, an IoT device — the threat model flips. The local user can read your binaries, intercept your traffic, modify your config, and lie to your control plane about what they are. The defenses are different from web-server hardening. Applies to: monitoring agents, RMM tools, CI/CD runners on customer infra, MDM agents, deployment agents, IoT controllers, sync clients, backup agents. ## When to invoke - Shipping a new native agent / endpoint client for the first time - Designing the installer or update channel for one - After an incident where a client was tampered with or impersonated - Auditing an agent product before integrating it (vendor review) - Reviewing an inherited agent codebase Pairs with [`distributed-system-audit`](../distributed-system-audit/SKILL.md) for the system view, [`ios-security`](../ios-security/SKILL.md) for mobile specifics, [`mcp-security`](../mcp-security/SKILL.md) for agent-like LLM tooling. ## The threat model — assume the host is hostile (eventually) For each thing your agent does, assume: - **The local user can read every file** the agent writes - **The local user can intercept every network call** (their own machine, their own CA store, their own packet capture) - **The local user can modify the agent binary** (replace, hook, debugger-attach, decompile) - **The local user can run multiple