Aleph

Security

Aleph combines inspectable bundles with isolated execution and scoped access. Security still begins with reviewing what an agent is instructed to do and granting only the configuration it needs.

Core boundaries

  • Bundle versions are immutable.
  • Turns run in isolated environments.
  • Secret values are encrypted and stored outside bundles.
  • Runtime values resolve from the agent's exact scope.
  • Access checks use ownership, visibility, organization permissions, and team membership.
  • Platform LLM access is managed securely by Aleph.

Declared values are made available only to the turn that needs them. Channel credentials remain separate from bundle files.

Review public agents

Public means discoverable and readable, not platform-endorsed. Inspect instructions, skills, hooks, schedules, and configuration requirements before cloning or enabling an agent.

Use narrow credentials, rotate exposed tokens, and never paste secrets into bundle files, conversations, or memory.

Supported extension boundary

Aleph supports bundle skills and sandbox hooks. MCP is not supported: do not add mcp.json, tool-server credentials, or MCP setup instructions to a bundle.

If you suspect unsafe behavior, disable the agent first. Disabling blocks new turns without destroying the evidence in versions or conversations.