Banks Shift AI-Agent Risk to Customers, Plus 3 Stories
Banks are drawing a bright line around AI agents and customer accounts, while two widely used pieces of internet infrastructure need security updates. The common thread is authority: who is allowed to act, which software is trusted to process outside input, and who owns the damage when those assumptions fail.
Banks say your AI agent may be your responsibility
When major banks began blocking outside AI agents last month, the practical question was whether these tools should be allowed into customer accounts. That story has now moved. Citi and PNC have rewritten their agreements to spell out what happens if a customer lets an agent in anyway.
Citi’s November Online User Agreement cautions customers against giving a third-party AI agent account access. If a customer provides credentials or otherwise permits access, Citi says it will consider the agent’s actions authorized, even when those actions conflict with the customer’s instructions. The agreement goes further: if the agent shares or leaks those credentials, Citi says later access using them will also be treated as authorized, except where applicable law says otherwise.
PNC’s updated agreement takes a slightly different route. It permits connections the bank specifically provides, with defined accounts, duration, and permissions. It prohibits giving an agent a password or one-time code, and it bars browser automation or screen scraping used to get around the bank’s controls. If a customer violates those restrictions, PNC says the customer is responsible for the resulting access and activity to the extent the law permits.
This isn’t a theoretical policy exercise. Google just made Gemini Tasks available to free users, with examples that include accessing a utility account through Google Password Manager, paying bills, shopping, and using payment cards. Google says sensitive actions require confirmation and the agent runs in an isolated virtual machine. Those are useful controls. They don’t turn a bank password into a safe integration method.
Now, the qualification the scarier summaries leave out: a bank agreement doesn’t automatically erase federal consumer protections. The Consumer Financial Protection Bureau’s Regulation E guidance says an agreement cannot impose more liability for an unauthorized electronic transfer than the regulation allows. If a third-party service is hacked and a thief moves money without actual authority, that may still be an unauthorized transfer. If you intentionally gave an agent authority and it exceeded your instructions, the dispute becomes much harder. Business accounts can also have fewer protections than consumer accounts.
That’s what bank AI agent liability looks like in practice: the line between “the software made a mistake” and “someone stole access” may determine who absorbs the loss. Don’t give an AI agent your banking password, passcode, or a shared browser session. Use a bank-approved connection, prefer read-only access, limit the accounts and duration, and review the agent’s existing permissions. If you’ve already shared credentials, change them, revoke active sessions, and inspect recent transactions. Convenience is useful. Ambiguous authority over your money is not.
This is a material update to our October brief on banks blocking unapproved AI agents. The banks aren’t merely blocking access now. They’re documenting who they intend to hold responsible.
Turning on AI is the easy part
Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.
NGINX fixes HTTP/2 memory corruption with public proof of concept
NGINX released stable version 1.30.7 and mainline version 1.31.9 on November 18 to fix several security problems. The lead issue, CVE-2026-123007, affects NGINX configurations accepting HTTP/2 traffic. F5’s advisory says crafted traffic can make an HTTP/2 flow-control value exceed its permitted limit, causing a heap buffer overflow.
Flow control is how HTTP/2 keeps one side of a connection from sending more data than the other side is ready to receive. Think of it as a running balance rather than a speed limit. The vulnerable calculation can overflow that balance and make NGINX write outside the memory reserved for it. The usual result is a crashed worker process. Researchers also demonstrated code execution when memory protections such as Address Space Layout Randomization are disabled or can be bypassed.
The NGINX advisory lists versions 1.9.5 through 1.31.8 as vulnerable to this HTTP/2 flaw. A public proof of concept exists, but CERT-EU says it has not seen exploitation in the wild. The same release fixes additional HTTP/2, HTTP/3, resolver, mail proxy, TLS, PROXY protocol, and stream-module problems. Disabling HTTP/2 can reduce exposure to two of the flaws, but it does not fix the whole set.
Administrators should update through the operating-system or product vendor, restart NGINX, and verify the running process loaded the fixed binary. Don’t decide from the version banner alone because Linux distributions may backport the patch without changing to NGINX’s upstream number. NGINX embedded in an appliance, Kubernetes ingress controller, OpenResty package, or hosting platform needs that vendor’s update. If a hosting company manages your web server, ask it to confirm the fix rather than trying to replace packages you don’t own.
That last distinction matters. Your WordPress dashboard cannot patch the web server underneath it. We covered that wider maintenance boundary in Your Website Is More Than the WordPress Dashboard.
WordPress security & maintenance
Keeping WordPress current is only part of keeping it healthy. Raymond Tec handles updates, backups, security monitoring, compatibility problems, access cleanup, and maintenance — plus the assorted weirdness that accumulates on a site over time.
Java4Shell is serious, but the hardware and application still matter
CISA is urging organizations to patch CVE-2026-114938, a memory-corruption flaw in the Base64 decoder used by affected Java Virtual Machines. Specially crafted Base64 data can crash the JVM and, under the right conditions, allow an unauthenticated attacker to run code.
Base64 is the common encoding that turns binary data into text safe for cookies, authentication headers, JSON Web Tokens, SAML messages, and other web traffic. The defect is in a hardware-accelerated decoding path called an intrinsic. It affects certain Java releases running on x86-64 processors with specific AVX-512 features, including newer Intel Xeon and AMD Zen server chips. ARM64 systems don’t use the vulnerable path.
CERT-EU’s advisory lists fixed OpenJDK releases including 17.0.21.1, 21.0.13.1, 25.0.5.1, and 27.0.1.1, but Java vendors use different build numbers and release schedules. Java can also be bundled inside an application, container image, or appliance where it won’t appear as a separately installed program.
Inventory first. Prioritize internet-facing Java services that decode attacker-controlled data, update to the fixed build from the JDK or product vendor, and restart every JVM. A downloaded JDK doesn’t replace the code already running in memory. Public scanners and proof-of-concept triggers exist, and GreyNoise has reported attempts, but successful exploitation in the wild has not been confirmed and the flaw is not currently in CISA’s Known Exploited Vulnerabilities catalog. Serious, yes. Universal, no.
Business IT goes well beyond the website
Your business also depends on workstations, cloud accounts, browsers, Wi-Fi, remote access, collaboration tools, and all the other technology that quietly becomes infrastructure. Raymond Tec works across that whole stack, whether the problem lives on a server, on a desk, or somewhere in between.
An EU legal opinion says personal data can be payment
An Advocate General at the Court of Justice of the European Union has concluded that a social network shouldn’t call itself “free” when access requires users to provide personal data and consent to its commercial use. The case began with Facebook’s old claim that the service “is and remains free.”
The November 19 opinion says commercial use of personal data has economic value and can count as payment under EU consumer-protection law. A service could still be described as free when it uses data only to provide the service or meet legal obligations.
This is a legal opinion, not the court’s final judgment. It doesn’t immediately change Facebook settings or create a new right for U.S. users. What it does is challenge a convenient bit of platform language. If the business receives something valuable in exchange for access, “free” may describe the lack of a credit-card charge without describing the actual bargain. The final ruling will determine whether that idea becomes law in this case.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources and further reading
- Citi Online User Agreement: third-party AI agents
- PNC Digital Services Agreement
- CFPB: Electronic Fund Transfers FAQs
- TechCrunch: banks update AI-agent terms
- NGINX security advisories
- F5: CVE-2026-123007 advisory
- CERT-EU: NGINX vulnerabilities
- CISA: critical Java vulnerability alert
- Oracle: November 2026 Java security update
- CERT-EU: Java4Shell affected and fixed versions
- Court of Justice of the European Union: Advocate General opinion
Photo by Polina Tankilevitch on Pexels.
