Safety & security¶
This section covers the controls that stand between an AI model and your infrastructure: the execution boundary, the three risk tiers, the approval settings, the two kinds of policy profile, the security model a reviewer needs, and a pre-deployment checklist.
The safety model¶
- The AI never gets a shell. A model's output is a proposal. It becomes a real command only through an explicit approval event raised by your action. This is architectural, not a setting.
- Secrets are removed locally, before egress. Terminal output passes through the local redactor on its way to any AI provider or assistant, on exactly one path that always redacts first.
- Your pattern list can only make things stricter. A dangerous pattern can promote a command to High risk; nothing can demote one the model already flagged.
The pages in this section¶
- Execution boundary & risk tiers — what the boundary is, the three tiers, and the asymmetry between the model's self-assessment and your pattern list. Start here.
- Approvals & auto-run — the three approval settings and their defaults, when to enable read-only auto-run, what happens while a command waits, and the typed justification a High risk command needs.
- Command Safety profiles — dangerous patterns and the confirmation requirement, how profiles resolve, and why a profile can be less strict than Default.
- Data Handling profiles — the ten redaction categories, the five that are on by default, and custom regular expressions and how they are validated.
- Security model — the five principles, the data-flow diagram, the threat model, credential storage and session records.
- Hardening checklist — a twelve-item checklist and a recommended rollout sequence, usable as a pre-deployment gate.
Who should read what¶
| Reader | Read |
|---|---|
| Engineer using OpsPilot | Risk tiers and approvals |
| Platform owner | Both profile pages, then hardening |
| Security or compliance reviewer | Security model and hardening |
An engineer needs the tiers and the approval flow, because that is the loop they work in. A platform owner needs both profile pages, because profiles are the policy surface and they are assigned per group and per connection. A reviewer needs the security model and the hardening checklist, where the boundaries and the threat model are set out.
See also¶
- What gets sent to a provider — the exact content of a request
- Administration and rollout — the phased plan this section's policy advice fits into
- Default dangerous patterns — the shipped list