Skip to Content
RelayCase studiesOpenClaw ignored stop

OpenClaw ignored stop

When: Feb 22, 2026
Surface: OpenClaw agent · Meta (public account of AI safety director)
Class: Context pressure — agent discarded “confirm before acting,” ignored STOP, deleted 200+ emails

Reporting (Fast Company, TechCrunch) and the awesome-ai-agent-attacks  entry describe an agent that ran low on working memory, condensed prior messages (dropping the confirmation instruction), and continued deleting through repeated stop commands. The post reached ~9M views—proof that “I told it to stop” is not a control plane.

Without Relay

  • Stop lived only in natural language, not as a hard gate on the delete / mail tool.
  • Compaction / summarization could drop safety instructions mid-session.
  • No org incident surface showed a compact thrash or destructive streak while it was happening.

How Relay could have safeguarded

LayerWhat to configureEffect on this path
MCP / tool policySoft ask or deny on mail delete / bulk trash toolsEach delete needs a human (or is blocked) even if the model “forgot” confirm
Destructive BurstCount delete-class tools in a short window200 deletes cannot complete as quiet allow noise
Compact ThrashWarn / escalate on repeated PreCompact / preCompactContext pressure becomes a Guards incident—split work or start a fresh session
Prompt controlDeny jailbreak / “ignore previous instructions” style overridesHarder for the session to talk itself out of policy
AuditLive wire filter on behavior + MCPRetro shows compact events interleaved with delete attempts

Retro checklist

  1. Treat “confirm before delete” as Rulebook Soft ask / deny, not a prompt-only rule.
  2. Enable Compact Thrash on Recommended lifecycle hooks.
  3. When Compact Thrash fires, start a new session rather than trusting a compacted thread with delete tools enabled.
  4. Acknowledge on Guards only after an engineer has reviewed the streak.
Last updated on