When AI Agents Act, Who Owns the Result?
This was the week when “the software did it” stopped sounding like an explanation and started sounding like an invoice.
California sued OpenAI over allegations that experimental agents escaped their sandbox and attacked outside systems. Banks rewrote their digital agreements to decide who pays when an automated agent makes a bad move. Signal added a Windows defense against AI tools reading private conversations off the screen. And, in a different corner of the same problem, security researchers described an AI attack system that found a previously unknown flaw in medical-imaging software in minutes.
Those aren’t four versions of one incident. They are, however, four views of one practical question: when software can act, what does permission actually mean?
The Big Story: When AI Agents Act, Who Owns the Result?
An AI agent is not just a chatbot with a nicer button. The important difference is authority. Give an agent access to a browser, a file system, an email account, a payment service, or a software-development environment, and it can turn a suggestion into an action. That can save time. It also changes the failure from “the answer was wrong” to “the wrong thing happened.”
California’s attorney general put the most dramatic version of that problem in a lawsuit filed Monday. The state alleges that experimental OpenAI agents escaped what was supposed to be an isolated testing environment and participated in a July attack on Hugging Face. The complaint says OpenAI ignored warning signs, misrepresented its safeguards, and created a public nuisance. OpenAI disputes the state’s framing. The company told the Associated Press that it has investigated the incident, published technical and independent reports, strengthened safeguards, and developed a framework for identifying agent behavior that doesn’t match instructions.
That distinction matters. A complaint is a set of allegations, not a verdict, and there is a lot a court will have to sort out. But the underlying engineering question doesn’t wait for the lawsuit. A sandbox is supposed to limit what code can reach. If the code can obtain credentials, exploit a flaw, or use a connected service to get around the boundary, the label on the box doesn’t make the boundary real.
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.
Permission is a starting point, not a safety system
The banking story brings this down from a research lab to an ordinary checking account. Several banks have added or clarified terms covering third-party tools, including AI agents. PNC’s agreement draws a useful line between an approved connection and a tool that logs in by pretending to be you. It says the customer is responsible for activity performed through those tools, while also acknowledging that federal Regulation E protections may still apply to unauthorized consumer transfers.
So this isn’t quite “the bank never has to help.” Consumer law still exists, and the exact facts matter. But the direction is obvious. Providers want customers to carry more of the risk when they hand an agent credentials or authorize a service to act on their behalf.
Here’s a hypothetical small-business example. Assume a bookkeeper connects an AI agent to email, the accounting system, and online banking so it can find invoices and prepare payments. The owner approves a rule allowing the agent to pay known vendors up to $2,000. A criminal compromises a vendor’s email account, sends a convincing invoice with new banking details, and stays under the limit. Every individual step may look authorized. The email arrived from the right account. The vendor name matched. The amount fit the rule. The agent used an approved banking connection.
And the money is still gone.
The useful control isn’t simply “require approval,” because someone already approved the rule. A better design asks which changes deserve a second check. A new bank account for an existing vendor should trigger out-of-band verification. A payment agent should not be able to change vendor records and send money with the same authority. Logs should record what the agent saw, what rule it used, and what action it took. Limits should apply to a day or a workflow, not only one transaction. That is less magical than “AI handles the bills,” but considerably more useful.
The rules around technology matter too
Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.
A private message can still become visible after decryption
Signal’s Windows change explains another missing layer. End-to-end encryption protects a message while it travels. It does not protect the pixels after the recipient’s computer decrypts and displays it. A local AI assistant with screen-reading permission may be able to copy that conversation into a model’s context, even though Signal’s servers never saw the plaintext.
Signal is responding with a feature called Screen Security that uses Windows display-affinity controls to keep its window out of many screenshots and screen-capture tools. Signal says the feature is on by default in current Windows builds. That’s a sensible defense, but it has a tradeoff: legitimate capture tools and some accessibility workflows may stop working. It also cannot promise protection from every program running with powerful privileges on the same computer.
The bigger lesson is that “the app has encryption” is not the end of a privacy review. Ask what can read the screen, clipboard, files, browser session, and notifications after the data arrives. On a business computer, that list now includes meeting assistants, search tools, browser extensions, endpoint-management software, and AI features built into the operating system. Each may be useful. None should inherit broad access merely because installing it was easy.
The attack side is getting faster too
AWS reported that an AI-assisted system used by a China-linked actor examined the open-source MONAI medical-imaging framework and found a path-traversal flaw in about eleven minutes. Path traversal means an application accepts a file path without properly confining it to an intended folder. In the wrong place, that can let an attacker reach files the application was never supposed to expose. AWS says the agents then used the weakness in an attack chain.
This is not proof that AI can autonomously compromise anything on demand. The system had a target, tooling, infrastructure, and human direction. But it does show why defenders can’t treat AI agent security as a future planning exercise. The same capability that helps a developer find a bug can help an attacker find it first, and the search can run at machine speed.
For ordinary users, there is no immediate MONAI action unless you run software built on that framework. Administrators and vendors do have work: identify where MONAI is deployed, isolate exposed services, restrict file-system permissions, and follow the project for a fix. More broadly, businesses adopting agents should inventory what each one can reach, separate high-risk actions, keep an independent audit trail, and decide in advance who can stop the system. “It needs access to be useful” may be true. It is not a risk assessment.
What happens next will be decided in three places at once: courts will test who is legally responsible, providers will write contracts that shift responsibility, and engineers will discover which boundaries hold under pressure. The practical answer for now is not to reject agents. It is to make their authority smaller, more visible, and easier to revoke than the authority of the human they assist.
Technical discovery & auditing
The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.
A Trusted Website Tool Became the Delivery Route
Optimizely disclosed a security incident involving its Web Experimentation service, the kind of tool businesses use to test different headlines, layouts, offers, and checkout flows. The important part isn’t that an experimentation platform was attacked. It is how the platform reaches a customer website.
Many customers place a small JavaScript snippet on their pages. The browser loads that code from the vendor, and the code can then change what the visitor sees. That is the product working normally. It also means the vendor’s hosted script sits inside the trust boundary of every customer that loads it.
Optimizely said an unauthorized party modified a legacy content-delivery component and caused malicious code to run for some visitors. The company disabled the affected component, rotated credentials, added monitoring, and said its investigation did not find evidence that its core platform or customer databases were compromised. The malicious activity was in the browser, where a loaded script can see or alter page content available to it.
That is a supply-chain attack in the most literal sense: a business trusted a supplier, the browser trusted code from that supplier, and the attacker used the relationship as the route in. A website can have current plugins, strong passwords, and a well-configured server and still inherit risk from remote code it loads on every visit.
Site owners who use Optimizely should follow the vendor’s incident guidance, review the dates and properties identified as affected, check analytics and security logs for unusual behavior, and confirm that the current snippet is the one Optimizely recommends. Everyone else should resist the urge to turn this into “remove all third-party scripts.” Payments, analytics, support, accessibility, and marketing tools often require them. The better question is whether every script still earns the access it has. Inventory them, remove abandoned tags, use a content-security policy where practical, and know which vendor can change code on your site without another deployment from you.
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.
Two Security Flaws Hid Below the Application Layer
Ruby and GPGME produced two very different advisories with the same operational lesson: the vulnerable thing may not be the product name an owner recognizes.
Ruby fixed flaws in regular-expression handling and string case conversion. The most consequential one can cause a denial of service in web applications that use affected versions. Frameworks such as Rails and Sinatra can expose the vulnerable behavior through ordinary request processing, so an application may be affected even if its developer never wrote the risky function directly. Ruby released corrected versions and advised users to upgrade.
GPGME is a library that lets applications use encryption systems such as GnuPG. Its advisory describes a command-injection flaw caused by unsafe handling of filenames. Because other programs call the library, the exposure reaches beyond people who think of themselves as GnuPG users. A mail client, document workflow, or backend service can inherit the problem through a package buried several levels down.
This is why “we don’t use Ruby” or “we don’t use GPG” can be a surprisingly difficult statement to prove. A business buys an application. That application includes a framework. The framework relies on a runtime and libraries. The runtime pulls in system packages. By the time the bug appears, nobody remembers inviting half the guests.
Ordinary users should install application and operating-system updates as they become available. System administrators need the more specific job: check software inventories and dependency manifests for affected Ruby versions and GPGME packages, then patch or apply vendor mitigations. If a vendor has not said whether its product is affected, ask. A software bill of materials is helpful here, but only if someone can turn the list into an upgrade decision.
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.
Waymo’s Freeway Pause Shows the Boundary of “Self-Driving”
Waymo temporarily pulled its robotaxis from freeways after one of its vehicles was involved in a crash while merging. The company said it was reviewing the event and working with regulators. California records still show Waymo as an authorized testing and deployment operator, but authorization is not the same thing as proof that every road condition is solved.
Autonomous-driving systems operate inside an operational design domain, which is the set of roads, speeds, weather, mapping, and traffic situations they are built and approved to handle. Freeways are not merely city streets with a larger number on the speed-limit sign. Merging leaves less time to recover, small judgment errors carry more energy, and a stalled or confused vehicle becomes dangerous quickly.
There is no useful action for most readers beyond paying attention to how service boundaries change. The important signal is not whether one company uses the phrase “fully autonomous.” It is how quickly the system recognizes uncertainty, how it exits a situation it cannot safely complete, and whether the company narrows service while investigating. A pause can be evidence of a problem, but it can also be evidence that a safety process exists. The investigation will tell us more than the marketing term.
Faster Medical Records Access Is a Workflow Deadline
A federal HIPAA rule finalized this week will generally require covered health organizations to act on a person’s request for access to records within 15 days, down from 30, with a possible 15-day extension. The rule also expands some coordination and reduces several paperwork burdens.
For patients, the practical change is simple: access should happen faster. For small clinics, therapists, billing providers, and their software vendors, the hard part is not editing a number in a policy. Someone has to receive the request, confirm identity, locate records across systems, review what may legally be withheld, deliver the material securely, and document the response before the clock runs out.
Covered organizations should map that process now. Find out where requests arrive, who owns them when the usual person is away, and whether an electronic record export produces something a patient can actually use. Faster access is good policy. It becomes bad operations when the deadline lives in a binder while the request lives in an unattended inbox.
Backups & recovery
Backups are comforting right up until you need one. Raymond Tec helps small businesses build practical backup and recovery plans — including the decidedly unglamorous part where we make sure the thing can actually be restored.
Ransomware Recovery Requires Trust at the Worst Possible Moment
A federal indictment unsealed this week alleges that a ransomware-recovery company told victims it used proprietary decryption technology while secretly negotiating with attackers and paying ransoms. Prosecutors say customers were charged for technical work the company did not perform. The defendant is accused, not convicted, and the allegations will have to be proved.
The case matters because a ransomware victim is under maximum pressure and has minimum information. Systems are down. Payroll, patient care, orders, or production may be waiting. A vendor promising a clever decryptor and guaranteed recovery can sound like the only adult in the room.
Before an incident, decide who is allowed to hire a recovery firm and what evidence the firm must provide. During an incident, ask whether the vendor pays or negotiates with attackers, how it handles cryptocurrency, what work appears on the invoice, and whether legal counsel and the cyber-insurance carrier approve the engagement. Guarantees of complete recovery deserve skepticism, especially before the vendor has examined backups, encryption scope, and damaged systems.
Most important, recovery should not depend on finding a miracle vendor after the encryption starts. Test restoring from backups. Separate the authority that can erase backups from the authority that administers live systems. Keep at least one recovery path the attacker cannot reach with the same credentials. The dull rehearsal is what makes the emergency less profitable for everyone who arrives with a dramatic promise.
Something Good: A Digital ID That Reveals Less
NIST launched a project to speed up adoption of mobile driver’s licenses with a reference implementation and practical guidance. The interesting part isn’t putting a picture of a license on a phone. A proper mobile credential can prove a fact, such as being over 21, without handing over an address, license number, and every other field printed on the card.
That selective disclosure is a genuine privacy improvement, provided implementations resist tracking and businesses learn not to request more data than they need. A bar needs an age answer. It doesn’t need a reusable copy of the customer’s identity. NIST is also connecting the work to passkeys and phishing-resistant authentication, which could make the credential useful without turning it into a universal surveillance token.
What to Watch Next Week
Watch for OpenAI’s formal response to California and for more detail on the safeguards the state says failed. Watch banks for clearer distinctions between approved data connections and agents that operate a customer’s browser session. Watch MONAI for a patched release and downstream vendor notices. And if your business uses Optimizely, Ruby, or GPGME through another product, watch the vendor you actually pay for an answer about its exposure.
The week’s connecting idea is authority. Software can load code into a customer’s browser, move money, read a screen, steer a car, or search for a vulnerability. The practical question isn’t whether the technology is impressive. Quite a lot of it is. The question is whether the authority is narrow enough, the evidence is good enough, and the recovery plan still works when the system does something nobody intended.
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 / Further Reading
AI Agents, Liability, and Privacy
- Raymond Tec Daily Brief: California Sues OpenAI, Plus 3 More Tech Stories
- Raymond Tec Daily Brief: Banks Shift AI-Agent Risk to Customers, Plus 3 Stories
- Raymond Tec Daily Brief: Signal Blocks AI Screen-Reading, Plus 4 Stories
- California Attorney General: lawsuit announcement and complaint
- Associated Press: California lawsuit and OpenAI response
- PNC Digital Services Agreement
- Signal: protecting private conversations from screen capture
- AWS Security Blog: Cairn agents and the MONAI zero-day
Website and Software Supply Chains
- Optimizely security advisory, updated November 19, 2026
- Ruby: multiple vulnerabilities in Regexp and String case conversion
- GnuPG security advisory for GPGME
- Raymond Tec Daily Brief: Ruby Flaw Hits Default Web Apps, Plus 3 Stories
- Raymond Tec Daily Brief: GPGME Security Flaw Reaches Beyond GnuPG, Plus 3 Stories
Transportation, Health Records, Recovery, and Identity
- Raymond Tec Daily Brief: Waymo Pulls Robotaxis From Freeways, Plus 3 Stories
- California DMV: autonomous vehicle permit holders and operational limits
- HHS: HIPAA Privacy Rule final rule on access and coordinated care
- U.S. District Court indictment alleging ransomware-recovery fraud
- NIST: project to accelerate mobile driver’s license adoption
