documentation / security

Security model

Ashplug reduces the risk of AI-assisted shell work by keeping generated content untrusted and routing every command through a centralized policy.

Default policy

allowpositively classified read-only commands.
approvalfile creation/modification, deletion, and network transfer.
denyprivileged operations, shell interpreters, unknown executables, and invalid commands.

Additional protections

  • Commands must be bare executable names, not paths.
  • Protected system and credential paths cannot be written or deleted.
  • Symlinks and write boundaries are resolved before policy decisions.
  • Timeouts and bounded output capture apply to every command.
  • Secrets and authorization headers are excluded from logs.

Limitations

The safety engine is not an OS sandbox. It cannot model runtime behavior, environment variables, plugins, shared libraries, or races between path checks and execution. Linux Bubblewrap isolation is opt-in and has its own documented limitations.