Skip to content

Identity-Level Kill Switch

Questions or comments?

Post them as a comment on the tracking issue -- requires a GitHub account with access to this repo.

Scope

In the event of a major violation, the agent's machine identity in ServiceNow can be disabled, cascading through provisioning to immediately halt all agent access and activity pending manual review. Per the source paper, this is the differentiator from governance frameworks that "stop at policy" — this one ends at enforcement.

Depends on ../agent-worker-identity-job-profile/spec.md (the identity to disable must exist) and ../human-supervision-escalation/spec.md (a major violation must be classified before this triggers). The exact cascade mechanics — "exactly how does an agent disable propagate to downstream systems" — are called out as a P2 TODO (Engineering, Not Started) in the source paper; this spec describes the intended outcome, not a confirmed implementation.

Value

Most governance frameworks stop at policy definition and audit; this one can actually enforce a stop, not just document that a violation happened. That's the concrete difference between an accountability framework and an accountability report.

Acceptance criteria

  • On a major violation escalation, an authorized human can disable the agent's machine identity in ServiceNow.
  • Disabling the identity halts all of that agent's access and activity — cascading through provisioning to every downstream system the identity had access to, not just the primary system the violation was detected in.
  • The halt is immediate, not eventually-consistent — no window where a disabled agent can still take a new action.
  • Reversing the kill switch (reinstating access after manual review) requires an explicit, auditable human action — it does not silently re-enable on its own.
  • The disable event itself is captured in the audit trail (who disabled it, when, tied to which violation).