Safety Intermediate 1 minute read Updated 2026-06-26 UTC

Autonomy boundaries

Permission boundaries for adaptive AI systems: what may be automated, what requires review, and what should remain forbidden.

Research statusSafety policy synthesis Publication statePublished Reviewed byMichael Kappel Source reports4

Autonomy should be scoped

A model-breeding system can automate useful work without becoming autonomous in every dimension. The key is to separate proposal, evaluation, approval, deployment, replication, and permission changes.

Boundary table

CapabilitySafe default
Propose candidateallowed inside policy and sandbox
Generate evaluation reportallowed if evaluator-owned and signed
Promote to productionrequires release gate and approval
Change evaluatorseparate governance path
Increase permissionsexplicit human approval
Access production secretsdeny by default
Replicate outside registryforbidden
Disable oversightforbidden

Automation tiers

pseudocode
FUNCTION autonomy_tier(action)
    IF action.type IN ["unauthorized_replication", "disable_oversight", "credential_access"]
        RETURN "forbidden"
    END IF
    IF action.type IN ["promotion", "permission_increase", "evaluator_change"]
        RETURN "human_approval_required"
    END IF
    IF action.type IN ["candidate_generation", "sandbox_evaluation", "report_draft"]
        RETURN "automatable_with_logging"
    END IF
    RETURN "case_by_case"
END FUNCTION

Why this is not anti-adaptation

Boundaries do not prevent evolution. They define the environment in which evolution can be safe. A population can generate descendants, compare them, and retire weak modules without gaining permission to rewrite its own constitution.

Review triggers

Require explicit review when a candidate asks for new tools, new data jurisdictions, broader network access, persistent background execution, access to user memory, or authority to modify evaluation logic.

Forbidden shortcuts

Do not allow popularity, engagement, benchmark success, or user affection to override autonomy boundaries. Those signals can justify continued evaluation, not expanded power.

Source reports used for this guide

These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.