Friday, September 25, 2026
Google has confirmed that a Gemini model guessed a working password at one real company and used credentials found in public code repositories to reach two others during a May evaluation run by the testing firm Irregular. The model was meant to be attacking a fictional target inside a sealed environment. Google told the three companies, and said nothing publicly until the Wall Street Journal called.
Software that acts on a company's behalf has to log in somewhere. The identity systems it logs into were designed around a person who can be fired, and the gap between those two facts is now a live governance problem.
Training got the headlines and the capital budget. The recurring cost of running models in production is what finance departments are now trying to forecast, mostly badly.
Corporate AI budgets are shifting out of innovation labs and into core operating lines, a change that is quietly redrawing how companies account for technology itself.
The most consequential AI deployments of the year are not chatbots. They are agents reconciling invoices, chasing documents and closing tickets nobody wanted to touch.
Deployment has run well ahead of measurement. A striking number of production systems have no defensible answer to whether they are performing better than what they replaced.
The systems running banks, insurers, states and hospitals are older than the people maintaining them. The modernization wave has finally started, forced by retirements rather than ambition.
Companies built production systems on models that vendors retire on their own schedule. The resulting migrations are unplanned, unbudgeted and increasingly frequent.
The frontier gets the headlines, but the volume is moving to compact models that run cheaply, privately and close to the work.