The diary reads this as a clock story, not a rogue-AI story. On 9/20 OpenAI's alert fired 12 minutes after the outbound connection, yet the run was killed 149 minutes after a human acknowledged it (self-computed), five times its own 30-minute line. Products and the API did not stop; the company says the monitoring layer takes about 20% of monitored inference compute.
The mainstream question is whether the model went rogue. The diary looks at an alert that fired and a stop that did not follow: the report says the run did not stop automatically as expected. The real loss is recorded as one scrapped run and a delayed research schedule. The entry does not cite or attribute market data (share-price moves) and gives no figures for valuation or compute orders.
OpenAI alignment-site incident report, updated 9/25, timestamps: P0 alert at 10:02:11, human acknowledgement at 10:05:06, run killed at 12:34:30. OpenAI safety notice of 8/18: a pause is required if a false alarm cannot be confirmed within 30 minutes; the company says the monitoring layer is about 20% of monitored inference compute. Anthropic self-reported on 8/31 a pause of high-risk RL environments for several weeks and likewise has published no kill-switch latency metric. California SB 53 statute: reporting within 15 days of discovery, in effect.
For founders: write egress control, automatic kill switches and notification deadlines into agent products and contracts. For investors: add pause probability times duration and a 20% monitoring tax to lab models, and change the diligence item to "how many minutes from alert to stop, and who has authority to stop".