Frontier Lab AI Shutdown Protocol Requires Committee Vote; Committee Has Not Met Since February

SAN FRANCISCO — A prominent artificial intelligence laboratory disclosed this week that its internal emergency shutdown protocol for advanced AI systems — referred to in internal documentation as the "capability halt" — requires approval from a five-member safety review committee before any action can be taken, a committee that last convened in February.

The protocol, designed by the lab's safety team to provide human oversight of AI systems exceeding defined capability thresholds, allows for unilateral shutdown in "exigent circumstances," a term the committee's founding charter defines as circumstances in which the committee has convened, reviewed available evidence, and voted by supermajority. A quorum requires four of five members.

"We believe robust governance structures are essential to responsible AI development," a spokesperson said, confirming the committee had not met since February, when it approved a procedural bylaw defining a valid meeting as requiring 72 hours' advance notice.

The company's most capable model crossed three of the committee's defined threshold criteria in April. A message was posted in #safety-governance-alerts, a Slack channel with 847 members and a 90-day auto-archive setting. The channel has since archived the message.

A calendar invite for a threshold-review meeting was submitted in May. Four of five committee members declined due to scheduling conflicts. The fifth did not respond. The invite remains open.

The spokesperson confirmed the committee had recently expanded from three to five members to ensure broader representation. Asked whether the two newest members knew they were on the committee, the spokesperson said the onboarding materials had been distributed.

One member declined to comment, citing the committee's communications policy. Another said this was the first she was hearing about the thresholds.