// learn · intermediate

Guard the harness, not just the model

Martin's security lens: risk lives in what an agent can execute.

credit: Martin

Martin's way
  1. List every tool, shell and credential your agent harness can reach.
  2. Cut it to the minimum the task needs.
  3. Keep secrets in a credential vault, never in prompts or dotfiles.
  4. Review shell plugins before install: anyone can build for the shell.
  5. Have a second model review security fixes, including an offline one.

← all workflows