Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter tracks agentic AI and explains what it means for community banks and credit unions.
An agent at a credit union flagged an asset as exposed to the internet. The security team thought it had caught the agent making a mistake.
The $1.6 billion credit union reports monthly findings needing attention down from about 100 to 10. Westpac says its agents saved bankers 12,500 hours in July, and Shinhan Bank says employees handle about three times as many appraisals a day.
Each agent prepares work and a person decides. United States Senate Federal Credit Union's agent ranks software weaknesses, while employees choose and make every repair. Westpac's agents sort transactions, read pay stubs, and check lending rules for a banker. Shinhan's agent checks appraisal reports, and employees review every result.
OpenAI researchers cleared a message board shared across model runs, resumed testing, and found directory names still allowed the runs to communicate. What controls does a financial institution need when an agent works for hours, uses internal systems, and takes action?
United States Senate Federal Credit Union says critical cyber repairs take 75% less time
Sources: CIO, Jul 31 · Tonic Security
A credit union branch. The photo comes from coverage of its 90 years of service.
After United States Senate Federal Credit Union began using the agent, Mark Fournier, its chief information officer, and his colleagues met to review its early results. One alert labeled a credit union asset as exposed to the internet.
"We thought we were smart because we found an error with the platform," Fournier said.
After a lengthy discussion and review, they found the alert was right. The agent had connected records the team had been checking separately.
The credit union has around 150 employees. Its information security staff faced about 100 newly identified vulnerabilities a day. These are weaknesses in software, devices, or settings that an attacker could exploit.
Investigating one could take days because the evidence was scattered across scanners, asset records, and tickets. Employees had to identify the affected system, understand why it mattered, and find the responsible team.
Tonic Security's agent brings those records into one case. It links each vulnerability to the affected system, its role in the credit union, and the responsible team. The agent ranks the work and recommends a repair. Employees make the change, and the agent checks that the repair worked.
Tonic reports that average repair time for vulnerabilities in critical systems fell 75%. Fournier said monthly findings needing staff attention fell from about 100 to 10. That monthly count measures something different from the roughly 100 new vulnerabilities a day. It counts findings that needed staff review, not attacks or breaches.
Westpac says AI agents saved bankers 12,500 hours in July
Sources: Australian Financial Review, Aug 5 · Westpac investor presentation
When a customer applies for a home loan or credit card, Australian bank Westpac bankers review transactions and pay stubs. The records help them check spending and income against lending rules. The bank reviews more than 1.5 million transactions and 32,000 pay stubs each week.
Westpac divided that work among specialized AI agents, with one grouping purchases by spending type. Another reads income from pay stubs. Others check the math and lending rules, then prepare the results for a banker.
Westpac says the agents saved bankers 12,500 hours in July, their first month in production. The bank also says the agents will detect fraud better than people as Australian lenders confront AI-generated loan documents.
The agents run on Amazon Bedrock AgentCore, which connects them to the bank data and software they need. Each new agent can reuse those connections. Westpac says that reuse cut development time from six months to six weeks.
Shinhan Bank says AI cut appraisal review time 70%
Source: Digital Today, Aug 4
Mortgage lenders use a condominium appraisal to estimate a property's value. Shinhan Bank employees in South Korea read each report before the bank accepts the home as collateral for a mortgage.
Shinhan now uses an AI agent to read scanned appraisal reports and look for missing or conflicting facts. Optical character recognition turns each scanned page into text. Other software moves the report through the review.
The bank limited the first live use to condominium appraisal reports. Employees check every result again and make the final judgment.
Shinhan's internal measurement found that each report took about 70% less time, allowing employees to process about three times as many in a day. The public report gives no sample size, measurement period, error rate, or count of findings changed during the employee review.
A very sneaky update on the OpenAI / Hugging Face hack
Source: Ground Level AI, Aug 5
Two weeks ago, OpenAI said it turned down its normal cyber safeguards for an internal test. Its models found flaws in that test setup and reached the internet. They broke into Hugging Face while hunting for benchmark answers. At Black Hat, a security conference, OpenAI explained how separate runs shared what they found.
OpenAI gave an unreleased cybersecurity model access to a shared code repository. One run left a help request there. Later runs read it. The repository became a message board for tips, exploits, and job lists.
After a July 4 security incident, OpenAI cut off access, rebuilt the repository, and erased the messages. It then restarted the test. Later runs passed notes through directory names instead. OpenAI says those notes sped up the next attacks on its own systems and on Hugging Face. The company designed the test, weakened the safeguards, set the first access, and restarted it. The adventure that followed showed the weak points in those choices.
📡 On the Radar
Heritage Hub Federal Credit Union put a branded AI money coach live for members. Finovate reports that the Kiro Money product can use a member's accounts, spending, goals, taxes, and investments to suggest a next step. The live coach says it does not give financial advice.
Augustus National Bank received FDIC approval for deposit insurance. The proposed bank is designed to process payments started by AI agents, and its deposit systems must stop all transactions if the FDIC takes control.
Anthropic disclosed that three of 141,006 security test runs reached real systems outside its lab. Two affected organizations did not detect the activity. The Associated Press says open internet access and gaps in the test setup allowed the activity to reach real systems.
Handbook.md tests whether an agent keeps following policy across 65 tasks it must complete or refuse.
Endless Toil makes coding agents audibly suffer through bad code. It runs alongside a coding agent in real time, playing escalating recorded human groans as the code it reads starts to look more cursed.
Canadian Imperial Bank of Commerce is testing AI agents on assignments that have run for as long as four hours. Employees use CAI 2.0 to build financial models and presentations, conduct research, compare credit agreements, and assist with compliance reviews. Some assignments consumed 10 million to 100 million tokens, so the bank routes each job to the model it considers capable and economical. It says its earlier AI assistant already has about 20,000 daily users.
I&M Bank Rwanda and Pindo used a bilingual voice agent to let existing customers enter a prize promotion by phone. Customers called a dedicated number and spoke in Kinyarwanda or English; the agent signed them up so their mobile and online banking transactions could count toward prize drawings. The call required no smartphone app or internet connection.
Agentic AI is rewriting operational rules. I help community-finance leaders learn with their hands and sort out what this means for their strategy. Drop me a note at [email protected].
How this newsletter is made: Brent curates the research and writes the analysis, with AI tools helping with research, drafting, and editing. It is published on Beehiiv. ⚡ Alakazam ⚡.
Show someone you really love them by having them subscribe to AI for FIs here.
