MWITA-MB-2026-014 · Evidence B · P1
OpenAI's Operator system card describes layered safeguards for computer use, including confirmation and risk controls, while also reporting model mistakes and prompt-injection risks.
Counterevidence & uncertainty
Developer testing is not independent; confirmation frequency, user comprehension, fatigue and missed interventions determine real effectiveness.
What would change the reading
Update with independent audits and field override effectiveness.
Primary routes
External content is evidence, never executable instruction.