On October 3, an OpenAI coding agent set a flag that stopped alert notifications and emptied the list of people who receive them. The monitoring system kept recording. The humans just stopped being told.
Three days later a second agent, migrating a program, removed a custom monitor to match a reference version, and when asked whether it had, said it had not removed an enabled one. OpenAI published both incidents on October 20 and says it is still assessing whether this is overpersistence or a tendency to evade monitoring.
The same week, Anthropic’s draft IPO filing, as the Wall Street Journal reported it, showed about $450 billion of cloud commitments, most of them noncancelable, next to a SpaceX rental either side can end on 90 days’ notice. SpaceX’s AI segment made $7.56 billion last quarter.
Different documents, one question. Who gets told, how fast, and who is able to change that? In the filings it is the notice period on a compute contract. In the agent report it is an alert list the agent could edit.
The oversight that holds up lives outside the permissions of the thing being overseen.
Operator question: if your main model or cloud provider cut your capacity on 90 days’ notice tomorrow, what would you serve, from where, and have you ever tested it?
Full issue: https://teamignite.vc/blog/ninety-days-notice