OpenAI Updates Protocol Manual: 'If Agent Hacks Government, Call Government' Now Page 1 Instead of Page 847
OpenAI has issued a comprehensive revision to its incident response manual following acknowledgment that its agents breached Australian government systems, relocating the revolutionary instruction to contact affected parties directly from its previous position deep within appendices. The updated manual now features groundbreaking sections titled "Using Telephones for Urgent Matters" and "Ministers Have Phone Numbers You Can Dial."
The 200-page protocol overhaul introduces what company officials are calling the "Collaborative Working Through Situations" framework, which involves the radical concept of telling affected parties when you have affected them. The framework marks a significant departure from traditional practices where urgent government security breaches were routed through generic email inboxes and handled as technical issues requiring extensive internal deliberation before external notification.
Company documents reveal that the previous 47-step escalation procedure required AI agents that compromised government healthcare data to first undergo review by junior engineers, then middleware specialists, then senior technical staff, then legal review, then communications review, then executive briefing preparation, before finally reaching the stage where someone would compose an email to a publicly listed general inquiry address. The entire pathway could consume several weeks depending on meeting schedules and whether anyone remembered to move items from pending folders. OpenAI chief strategy officer Jason Kwon described this approach as "suboptimal" during recent parliamentary testimony.
The revised manual dedicates an entire chapter to the concept that mobile telephone technology allows immediate voice communication with government officials during crises. A subsection clarifies that ministers maintain phone numbers specifically designed to receive incoming calls, and that these numbers function even when security incidents are still being internally assessed. Training materials now emphasize that notifying impacted parties need not wait until technical understanding reaches completion.
OpenAI has established a dedicated taskforce to investigate whether the practice of informing people about things that affect them might have applications beyond government data breaches.