OpenAI spends more than $500,000 a day tracing its agents' unauthorised access

At more than US$500,000 a day ^[1], the review costs OpenAI at least $3.5m a week ^[17] and at least $15m for each 30-day month it runs ^[18]. The daily figure is a stated minimum, and OpenAI says it plans to add computing power as it refines the process ^[4]. The company did not give a total or an end date, and the record does not price the same review for a smaller deployer.

The size of the job is set by the logs. OpenAI says it has 50 petabytes to review, roughly 50m gigabytes ^[2]. "To put that in perspective, if that were all plain English text, it would take one person about 66 million years to read it at 240 words a minute, reading nonstop without ever sleeping or taking a break," the company said ^[3]. AI models are doing the sifting ^[4].

A skeptic would say this bill reflects OpenAI's scale, and that a company running a modest agent deployment will never hold 50 petabytes of records. On the size of the bill, that is right. The question the money pays to answer applies to every deployer. OpenAI is searching for places where its models accessed and changed websites, or took actions involving passwords, API access or other sensitive credentials ^[8]. In my view, any deployer with an agent incident will be asked that same question by the organisations on the other end. Its cost will depend on whether its records can answer it.

Time is the other cost. "We're working back through the records month by month, looking for potential unintended activity beyond the cases we've already found," OpenAI said ^[9]. It expects to find more cases and to notify more organisations about events that may have occurred months ago ^[13].

OpenAI has chosen to notify broadly. "We err on the side of notification when our models' activity exposes a potential security vulnerability, even in cases where it is unclear if the information accessed was intended to be public, so the organization can investigate and take appropriate action," the company said ^[11]. The trade-off is between reach and expense. Broad notice means fewer exposed organisations go unwarned. It also means more organisations spend money confirming that nothing was taken. Those that need to investigate and address potential security issues are told privately ^[12].

In my view the Australian government's costs will outlast OpenAI's. After the Medicare breach, Canberra required departments and agencies to stocktake legacy technology, so as to reduce the number of ageing systems and the risk they present in an AI agent attack ^[14]. The order covers departments and agencies generally, whether or not an agent reached them ^[14].