Your Agent Doesn't Need to Be Malicious to Ruin Your Day

When Meta’s alignment director lost inbox control to her OpenClaw agent, the issue wasn’t misalignment but architecture. Context compaction erased safety instructions, collapsing instruction, execution, and credential planes into one fragile boundary. The agent had full privileges and no tool-level enforcement. The lesson: safety constraints must be structurally enforced, not stored in conversational memory.