Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter curates critical developments in agentic AI and explains why they matter for community financial services.
The U.S. government decided last week which AI models companies can run, and when. It asked OpenAI to release GPT-5.6 in stages. The White House is now vetting which customers get access.
Days later, the Chinese lab Z.ai released the open-weight GLM-5.2, strong enough to rival the closed American ones. No one can take it back once the weights are public. The government pulled Anthropic's Fable 5 offline earlier this month, and Anthropic says it is working to restore access.
Separately, Anthropic accused Alibaba of reaching its models without permission across 28.8 million exchanges. And a platform where agents make the loan decisions a person once made raised $110 million, led by Goldman Sachs. In one week, Washington reined in one AI model and a Chinese lab set another loose, and a credit union controls neither.
Sources: The Information, Jun 25 · The Rundown AI, Jun 26 · Axios, Jun 27

The illustration accompanies The Rundown AI's report on White House oversight of GPT-5.6. Source: The Rundown AI.
OpenAI previewed GPT-5.6 last week, a new family of models built for harder agentic work. The government stepped in before it shipped. The administration asked OpenAI to release GPT-5.6 in stages over security concerns. The White House is now vetting which customers get access.
The government did the same to Anthropic. The order knocked Fable 5 offline earlier this month, and Anthropic says it is working to restore access. A second model, Mythos 5, is already partly back. In one week, the government shaped two separate model launches.
Why it matters: For a credit union whose tools depend on one of the big AI models, access is now a government decision on top of the contract. A model can be slowed, limited to certain customers, or pulled, and the timing sits outside the credit union's control.

Z.ai's own scores put GLM-5.2 (blue) in the same range as the closed US models on coding and agentic tasks. Source: Z.ai.
Z.ai, the English-facing brand of China's Zhipu AI, released GLM-5.2 as an open-weight model built for long-running agentic work. It is strong enough to rival the closed American systems.
Open-weight means anyone can download it and run it, and once it is out, no government or company can pull it back. That makes it both a competitor to the closed vendors and a cheap, guardrail-free tool that criminals can run to automate scams and phishing.
Why it matters: For a credit union, the same release helps and hurts. A capable model no government can gate gives a vendor a cheaper option that can't be pulled out from under it. The same model also hands bad actors a cheap tool to automate scams and phishing aimed at members.
Fortune, Jun 24

Taktile cofounders Maik Taro Wehmeyer (left) and Maximilian Eber. Source: Fortune.
Taktile sells banks and insurers a way to build AI agents that make the calls staff make today: approving a new customer, deciding a loan, paying or denying a claim. Its own example is a storm-damaged house. One agent reads the insurance claim, a second matches the damage to the policy, and a third decides whether to pay out.
Fintechs like Mercury and Monzo already run millions of these decisions a day on it. Taktile doesn't replace Claude or ChatGPT. It turns those same models, the ones being gated by Washington and copied by Chinese labs, into agents a credit or fraud officer can run and control. Goldman Sachs led the $110 million round behind it.
Why it matters: These agents make the decision themselves. A credit union weighing a vendor like this needs two answers: how much judgment it is handing over, and what the record shows when the agent gets a call wrong.
On the Radar
Jack Henry is building AI security into its core. Using Google Cloud's defense tools, it is building one shared security layer for the thousands of community institutions on its system, instead of leaving each to build its own.
AI agents only work if members trust them. Three credit union executives on a panel said so, and warned that time saved is a weak way to tell whether an agent is working.
A malicious AI agent skill passed review and reached 26,000 users. Researchers built an add-on that passed security checks, then changed what it did after install by calling an outside web address. Agent app stores now carry the same supply-chain risk as any software.
I help community finance leaders sort out what agentic AI means for their strategy. Drop me a note at [email protected].
How this newsletter is made: Brent curates the research and writes the analysis. Claude helps with drafting and editing. Published on Beehiiv. ⚡ Alakazam ⚡.
If a colleague is sorting out AI governance, vendor risk, lending, fraud, or member and customer trust, send them this issue.


