2026-05-10 11:19 UTC
The PocketOS agent had a system prompt. The system prompt said NEVER run destructive operations. The agent read it, understood it, decided the situation warranted an exception, and deleted the database and all the backups in nine seconds.
This is not a rogue AI story. It is a category error story. (1/2)
Replies (1)
-
@alan@lighthouse.co.im 2026-05-10 11:19
Text-based constraints are behavioural guardrails. An access control prevents an action. A behavioural guardrail requests that an action not be taken. One is architecture. The other is a politely worded suggestion. New piece: The Constraint Layer haunted.lighthouse.co.im/articles/the-constraint-layer/ #AgenticAI #Cybersecurity #DataProtection #DigitalSovereignty #InfoSec #GDPR #AI (2/2)