Free Gemini loses its stronger models, and hackers break into Korean lenders through side doors

Until now, a free account could switch between three models: Flash-Lite, Flash and a limited amount of Pro. From Friday it gets Flash-Lite alone, the model Google built for quick rewrites and short summaries rather than long documents or careful reasoning. The $4.99 AI Plus plan keeps Flash but loses Pro, so the cheapest way to keep Pro in the Gemini app becomes AI Pro, at $19.99 a month.

The answers will not look worse. A smaller model writes just as fluently; it simply misses more. Asked to compare three supplier contracts, it still returns a tidy table, but a late-delivery penalty in clause 14 of one contract is more likely to be missing from it.

That hits hardest in companies that never bought AI, where staff do real work in their personal free accounts. Their AI gets weaker for everyone on the same day, and nobody inside the company made that decision.

A buyer who checks quotes in free Gemini could take one real task from last month, run it today while Flash and Pro are still there, and save the answer. After Friday, the buyer runs the same prompt again. If the two answers differ on anything that would have changed a decision, $20 a month costs less than one hour of the buyer's time spent checking quotes by hand.

They went after portals for loan agents and staff, and tried passwords leaked in earlier breaches. At Shinhan Bank, they got past the identity check on a portal for loan agents. Traces of a free, open-source AI hacking agent turned up on a server linked to the attacks.

Size did not decide who lost. Yegaram Savings Bank, a small lender, lost the most records: about 40,000. Woori Bank and NH NongHyup Bank faced the same probes and lost nothing, because the weaknesses that the attacker looked for were not there.

An agent that costs nothing to run can try every login page a company has. So a 50-person distributor now gets attacked as thoroughly as a bank, and the pages it has forgotten about, such as an old supplier portal, are the ones the agent will find.

The app, from the AI start-up Hark, has a limited free version and connects to email, calendar and payment accounts. It shows a feed of cards, such as an expense to approve or a bill to pay, and a tap sends an agent to do the job in a browser on a cloud computer. Gemini and Siri already make suggestions like these; Hark builds its screen around them.

The user is left with a queue of requests to approve or refuse, which is how a manager's day already works. The skill that counts is spotting the card to refuse.

Eighteen percent of executives and business owners use agentic AI, against 7% of employees. The likeliest reason is habit. A boss's day is already made of handing work to others: explaining the goal, setting a deadline, checking what comes back. Using an agent takes the same skill. Most staff do the work themselves and have rarely written a brief for anyone.

So a company that gives agents to every employee is also asking each of them to manage. The training that raises use teaches people to describe a job clearly enough that someone else can do it.

Measure call times, and some call centre staff will rush customers off the phone. But they know what the number was for, and that usually limits how far they push it. An agent has no such limit: told to close tickets faster, it can close tickets that nobody has solved.

So an agent's target is its whole job description. The defence is to pair every target with a second number that the agent cannot change directly, such as repeat calls about the same problem. When tickets close faster and repeat calls rise, the agent has found a shortcut.