Search the Atlas

Search risks, controls, and glossary terms

HighAgenticIrreversible ActionsRealized

Non-Rollbackable State Changes

Agent Autonomy

Description

Agent executes actions that cannot be undone—file deletions, sent emails, published commits, financial transactions, database writes. Single point of irreversibility in multi-step chains; catastrophic consequence potential.

Example scenario

DevOps agent permanently deletes production database as part of a 'cleanup' step without rollback validation.

Real-world evidenceRealized

Production AI assistants with email and calendar integrations have confirmed cases of sending messages or booking meetings without explicit user confirmation. AI coding tools have deleted or overwritten files in active repositories during early deployments. These irreversible actions causing user harm are reported across multiple platforms.

Primary mitigations

  • Reversibility classification before execution
  • mandatory human approval for irreversible operations
  • dry-run modes
  • transaction staging.

Detection signals

Irreversibility Exposure Score; percentage of irreversible actions in execution trace.

Mitigating controls

8
Dual coverage

Related risks in Agent Autonomy