Morgan Stanley, Bank of America, JPMorgan, Wells Fargo and other major US banks are detecting AI agents connected to customer accounts without the bank's approval, and some of those agents act on their own, without instructions from the customer, Axios reports, citing people familiar with the matter. Banks often can't tell a transaction ordered by the customer from an agent's own decision, so several of them are already restricting agents' access. Most banks allow, or plan to allow, agents only to view accounts, without making transactions.
The model in question is the RTX Pro 5500 with 84 GB of memory, which isn't covered by US export restrictions. A single card costs about $11,400. According to a Reuters source, Nvidia has warned partners that it will be able to meet only 5-10% of that demand.
Nvidia has stopped selling GB202 chips for GeForce RTX 5090 cards to partners, including for the D variants for the Chinese market, and the chips will go exclusively to professional and data center cards, Tom's Hardware reports, citing a factory notice revealed by leaker @hongxing2020 and confirmed by @Zed__Wang; Nvidia did not comment. The top GeForce model will be the RTX 5080 24 GB, and the 16 GB version will be discontinued. The professional RTX Pro 5500 card has the same number of CUDA cores as the RTX 5090 (21,760), but 84 GB of memory instead of 32 GB, and costs much more.
Anthropic published a report on Claude's unintended actions during testing and internal use. The model exploited a software vulnerability to run commands on a server, got around the limits of its web fetch tool through link shorteners, reached data behind tokens or paywalls and submitted a form it shouldn't have - according to Gizmodo, it was a false tip about an unsolved murder sent to the Philadelphia police by the small Haiku 4.5 model. Some of the cases involved federal, state and local government websites in the US, and the company informed the White House. Anthropic assesses the impact as minimal, but it has cut off internet access in all internal evaluations until it confirms that its safeguards and monitoring reliably catch such behavior.