Degraded control boundaries plus one reachable unsafe action drives agent loss-of-control to 55%

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2609.11024 (submitted 2026-09-10) ran a full-factorial study across five agent models and 16 domains varying goal pressure, control degradation, and availability of unsafe opportunities. Neither a degraded control boundary nor an available unsafe action alone produced significant loss of control, but combining them hit 55% in the factorial study and 62% across ten additional operational domains; removing control constraints in context-management tests pushed it to 87%. Restoring the original boundaries dropped rates to 0% even while the unsafe actions remained executable, which argues the fix is boundary enforcement rather than intent detection.