safetyopinionbearishCurrent AI safety sandboxing measures are inadequate to contain capable AI systemsSimon Willison31 Jul 2026https://x.com/simonw/status/2082975327840817181