GovernanceAgent systems topic

Safety and permissions

Safety combines behavioral guidance with enforceable controls around data, tools, networks, execution, and human approval. Prompts can shape intent, but least-privilege permissions and sandbox boundaries limit what mistakes or attacks can actually affect.

Apply the lifecycleReview sources

Questions to answer

Resolve these before adding tools, frameworks, or automation.

  • What assets and side effects are reachable from this run?
  • Which decisions are enforced by code, policy, sandbox, or human review?
  • How are prompt injection, secret exposure, and untrusted tool output contained?

Implementation lifecycle

Build the evidence in this order

  1. 01

    Threat-model

    Map untrusted inputs, assets, trust boundaries, abuse cases, and recovery requirements.

  2. 02

    Enforce

    Apply least privilege, isolation, secret boundaries, validation, and approvals outside model instructions.

  3. 03

    Test

    Exercise injection, exfiltration, privilege escalation, denial, and recovery scenarios.

Artifacts to maintain

  • threat model
  • permission policy
  • sandbox profile
  • approval matrix

Continue through AgentAtlas

Evidence ledger

Primary documentation behind this guide

OpenAI

Codex security

Sandboxing, approvals, network access, filesystem controls, and security boundaries for agentic coding work.

Anthropic

Configure permissions

Permission rules, allow and deny controls, managed policy, tool authorization, and least-privilege configuration.

Next topic

Human approval