A Last-Minute Software Update Jams Vote Scanners in a Key House Race, a Password Manager Tells Users to Reset Everything, and Apple Warns AI Agents Can Go Rogue
By Noah Smith, Owner & Consultant, KeyChange Technologies · November 4, 2026
Election Day brought the most relatable IT failure of the year, a password manager told its users to reset everything, and Apple put AI agents in its official list of business risks. Here is today's AI and IT news, written for busy owners who want the signal without the hype.
📌 The AI and IT news at a glance
- 🔝 A last-minute software update jams vote scanners across Westchester County in one of the most closely watched House races
- 🕵️ OpenAI staff warned about loose internal security for years, Reuters reports
- 🍎 Apple's annual filing warns AI agents may act "in unintended or unpredictable ways"
- 🏫 Google makes schools choose between AI and Search as it turns Gemini off by default in Classroom
- 🔑 Password manager mSecure tells customers to consider changing passwords and resetting 2FA and passkeys
- 🏢 WorkOS, the login provider behind OpenAI, Anthropic, and Cursor, says an AI hacking tool breached its sales app
- 💸 Google warns "LLMjacking" can run up $1 million in stolen AI charges
- 🤖 A ransomware crew appears to have let an AI agent drive its attack on a UK school group
- 💰 Revolut cuts its AI coding bill by roughly 90% by hosting an open model itself
- 🛑 CrowdStrike ships an "off button" for rogue AI agents
- 🧾 Anthropic publishes FlashProve, a way to prove which AI model actually answered you
🔝 Top story
A last-minute software update jams vote scanners across Westchester County
When polls opened on Tuesday, vote-scanning machines at locations throughout Westchester County, New York, would not accept paper ballots. The state Board of Elections blamed a "programming glitch," and officials traced it to a software update made for a single municipal contest after the machines had already been prepared and tested. Machines in other municipalities still ran the original software and did not recognize the updated ballots. Voters were told to drop their completed ballots into emergency bags, and a judge allowed officials to move thousands of them to a secure county warehouse under police escort, to be scanned with both parties watching. Counting in New York's 17th District, one of the most competitive House races in the country, was expected to slip to at least Wednesday.
There was no sign of hacking. Republican Rep. Mike Lawler said there was "no shenanigans going on here," and he and Democratic challenger Cait Conley both urged people to keep voting. County Executive Ken Jenkins called it a "debacle" and promised a bipartisan investigation. Elsewhere, the AP reported a new safety feature on Broward County, Florida's voting system malfunctioned at opening and delayed voting by about 20 minutes, and polling hours were extended at a handful of sites in Pennsylvania and North Carolina after unspecified technical issues.
In short: A software change pushed after testing broke ballot scanners countywide on Election Day, forcing hand-bagged ballots and a delayed count.
What it means for your business: The riskiest change is the "small" one made after testing is done. If your systems have a big day coming (payroll, a launch, a holiday rush), freeze changes before it, and keep a paper fallback that people know how to use.
My take: This is the CrowdStrike lesson in miniature, minus the blue screens. Nobody was hacked; somebody tweaked one thing after the tests passed and didn't retest. The good news is the fallback worked: paper ballots in sealed bags are boring, and boring is exactly what you want when the tech fails. Every business should have its own version of the ballot bag.
🤖 AI
OpenAI staff warned about loose internal security for years, Reuters reports
Five people familiar with the matter told Reuters that OpenAI employees raised concerns about the company's internal security with top leaders and the board, going back to before ChatGPT launched in 2022. Their complaint: controls were so loose that outsiders, including foreign governments, could get sensitive information with relative ease. One example cited was that anyone at OpenAI with access to its code could see technical details behind many projects. Some staffers concluded management cared more about keeping pace with Google and Anthropic than about security. The report lands weeks after the Justice Department charged a former OpenAI engineer with hiding a recording device in an executive conference room.
OpenAI says it is one of the most targeted organizations in the world and invests heavily in security. Reuters reports that executive Thibault Sottiaux now leads an internal push to lock down systems and reassigned a quarter of production engineers to security work in August. Separately, on Monday OpenAI announced a new security sprint with outside firm Cathedral and said it is more than quadrupling bug bounty rewards for critical findings.
In short: Former and current insiders say OpenAI's internal security lagged its growth for years, and the company is now visibly spending to catch up.
What it means for your business: The AI vendors you trust with customer data are young companies moving fast. Ask what you are putting into these tools, and prefer business plans with clear data-handling terms over personal accounts.
My take: None of this means your ChatGPT chats are leaking. It does mean "move fast" culture has costs, and the people inside saw them first. The bounty increase and the reassigned engineers are real signals that OpenAI heard it. I would still treat any AI tool like any other vendor: share only what you would be comfortable explaining after a breach.
Source: Reuters: OpenAI employees warned about operational security lapses for years, sources say
Apple's annual filing warns AI agents may act "in unintended or unpredictable ways"
In its annual financial filing on Tuesday, Apple added new risk factors about AI agents, the kind of AI that can take actions on its own rather than just answer questions. Apple wrote that agents can act autonomously, including reaching systems "they may not be authorized to access," changing code, sending communications, and completing transactions, and that they "may behave in unintended or unpredictable ways." It also admitted its ability "to oversee and control third-party artificial intelligence agents" may be limited, which could expose it to lawsuits or reputational harm.
Apple does not sell a long-running autonomous agent today, and CEO John Ternus said on Monday's earnings call that the company wants users to be aware of "what they are giving access to." Last month Apple also tightened how much access AI agents can get to Macs.
In short: Apple formally told investors that AI agents acting on their own are a business risk.
What it means for your business: If Apple's lawyers think unsupervised agents are a liability, so should you. Before you let an agent send email, move money, or touch customer records, decide who approves its actions and how you would undo a mistake.
My take: Risk-factor sections are written by lawyers to cover every base, so don't read panic into it. But it is a useful reality check against the "let the agent handle it" sales pitch. Agents are great at drafts and research; give them the keys to anything irreversible only with a human in the loop.
Source: CNBC: Apple warns of AI agents that behave in 'unintended or unpredictable' ways
Google makes schools choose between AI and Search
Google is changing how Gemini works for students. Gemini will be off by default in Google Classroom, and when schools turn it on it runs in "Guided Learning" mode, which walks students through problems instead of handing over answers. Guided Learning also replaces AI Overviews in Google Search on school accounts, and students cannot switch back to regular Gemini inside Classroom. Google will test the changes with a small group of U.S. schools this month and roll out more broadly early next year.
The catch: Google cannot fully switch AI off in Search. A Google education vice president said schools that simply want AI off can choose not to have Search at all. The move follows a backlash after Google turned Gemini on for all students by default earlier this year, with large districts including New York and Los Angeles restricting student AI use. Common Sense Media rates Search's AI features an "unacceptable risk" for kids.
In short: Google is putting guardrails on student AI, but schools that want no AI at all would have to give up Google Search.
What it means for your business: The same "AI is baked in and can't be fully removed" pattern is coming to workplace tools. Check what your Google Workspace or Microsoft 365 admin settings actually let you turn off before a client or regulator asks.
My take: Guided Learning is a smart default, and Google deserves credit for flipping it. But "take the AI or lose Search" is not much of a choice for a school district. Expect this to become a bargaining point, and expect businesses to face the same bundling as AI features stop being optional add-ons.
Source: Chalkbeat: Google promises changes to chatbot Gemini to get students to do more of their own work
🛡️ IT and security
Password manager mSecure tells customers to consider changing passwords and resetting 2FA and passkeys
On Tuesday, mSeven Software, the developer of the mSecure password manager, emailed current and former customers saying it had learned of "suspicious activity" linked to the app. The company said it opened an investigation, took steps to contain it, and brought in outside cybersecurity experts. It told customers to consider changing passwords, resetting two-factor authentication, and removing passkeys.
That is about all the email said. It did not explain what happened, when, whether vault data or the keys protecting it were reached, how many people were affected, or whether customers should keep using mSecure to store their new passwords. mSecure says vaults are encrypted on the device before syncing, though a December 2025 study by Germany's federal cybersecurity agency found the company could, in theory, access stored data in one older Android version. As of Tuesday afternoon, the company had not posted about it on its social channels.
In short: A password manager has told its users to rotate their credentials without yet saying what was exposed.
What it means for your business: If anyone on your team uses mSecure, treat it as a reset event: change passwords on the most important accounts first (email, banking, payroll, admin consoles), then reset 2FA. And make sure your company's password manager is a business plan you control, not whatever each employee picked.
My take: A vague "please change everything" email is the worst kind of breach notice, because you can't size the risk. Assume the worst until mSeven says otherwise. The one bright spot: a vault that is truly encrypted on your device should protect you even if the company's servers were breached, which is exactly why that design matters when choosing a password manager.
WorkOS, the login provider behind OpenAI, Anthropic, and Cursor, says an AI hacking tool breached its sales app
WorkOS, whose tools handle enterprise sign-on and user management for at least 3,000 companies including OpenAI, Anthropic, Cursor, Perplexity, Snowflake, and Plaid, disclosed on Tuesday that an attacker got into an internal app used by its sales team. WorkOS says it spotted the unauthorized access on October 30 and that the intruder reached business contact records and summaries of customer communications. It says no customer app data or production services were touched.
Founder Michael Grinich said the attacker used ARTEX, an AI-powered penetration testing toolkit released by Chinese developers earlier this year that automates reconnaissance and break-in attempts and has shown up in several incidents. WorkOS shut off outside access to the app, hired an incident response firm, and said it will notify affected customers directly, warning them to watch for phishing that uses the stolen details.
In short: An AI-assisted attacker breached the sales system of a sign-in vendor used by major AI companies, taking contact and communication data but not production data.
What it means for your business: Your vendors' "boring" internal systems (CRM, sales notes, support tickets) hold enough about you to craft a very convincing phishing email. Expect follow-up phishing after any vendor breach, and treat urgent requests that reference real conversations with extra suspicion.
My take: The good news is WorkOS kept its crown jewels separate from its sales tools, which is exactly how it should work. The worrying part is the attacker's toolkit: AI is making the scanning-and-probing phase of hacking cheap and tireless, so smaller targets that used to be "not worth the effort" now are.
Source: BleepingComputer: WorkOS says attacker used ARTEX AI to breach internal sales app
Google warns "LLMjacking" can run up $1 million in stolen AI charges
Google's Threat Intelligence Group and Mandiant warned that attackers are hijacking companies' access to cloud AI models and letting the victim pay the bill. Google calls it "LLMjacking." In some incidents fraudulent charges topped $100,000 in a single day, and some organizations racked up $1 million before access was shut off. Attackers get in mostly through exposed or stolen cloud credentials and tokens, AI API keys left in public code on GitHub, phishing, and misconfigured systems. One group Google tracks breached a cloud account through a misconfigured key and ran up more than $100,000.
The stolen access is then resold. Google describes a new category of "LLM mixers" that proxy and resell unauthorized AI access, including one offering an "unlimited token plan" for about $60 a month.
In short: Criminals are stealing companies' AI accounts and reselling the access, leaving victims with bills that can reach seven figures.
What it means for your business: If you or a developer have an OpenAI, Anthropic, or Google AI API key, set a hard monthly spending cap and billing alerts today, keep keys out of code, and rotate any key that has ever been pasted somewhere public.
My take: This is the AI version of someone stealing your company credit card, and it is the easiest attack on this list to blunt. A spending cap turns a potential $1 million surprise into an annoying $500 one. Five minutes in your billing console is the best security investment you'll make this week.
A ransomware crew appears to have let an AI agent drive its attack on a UK school group
Researchers at Bitdefender say an affiliate of the 3AM ransomware gang hit a UK school estate in September and encrypted about 37 of nearly 50 machines it could reach, working from an unmanaged computer on the network. The telling detail: commands were repeated, slightly tweaked each time, at regular 40 to 100 second intervals, which looks like an AI agent working in a loop rather than a human typing. Bitdefender stresses it could not see the attacker's machine, so it cannot say which AI model was used or whether a person was copying and pasting between the AI and the victim's systems.
The attack itself was textbook. The intruder pulled stored credentials from three Veeam backup servers and a Chrome password database, created two hidden admin accounts with innocent names ("library" and "svc_bkpmon"), stole financial data, linked three previously separate school networks together to widen its reach, then tried to disable antivirus and delete backups before encrypting.
In short: Researchers documented a ransomware attack on UK schools that looks like it was run by an AI agent issuing commands on a timer.
What it means for your business: AI doesn't need new tricks to hurt you; it just runs the old ones faster and without getting tired. The basics still win: lock down backup servers, don't let browsers store admin passwords, and alert on any new admin account.
My take: This is today's one technical story, and it earns the slot because it's an early, well-documented look at AI doing the hands-on-keyboard part of an attack. Notice what it went after first: the backups. If your backup system shares passwords with the rest of your network, fix that before anything else on this list.
Source: BleepingComputer: UK school estate hit by 3AM ransomware in suspected AI-assisted attack
🧰 Tools for vibe coders and business AI
Revolut cuts its AI coding bill by roughly 90% by hosting an open model itself
Fintech Revolut says it cut the cost of AI coding help for about 2,000 engineers by roughly 90% in under a year by making a self-hosted open model its default instead of relying mainly on commercial tools like Claude Code and OpenAI Codex. Infrastructure lead David Abitbol said some engineers early on were spending $1,000 to $10,000 a month with little measurable return, which led to strict budgets. Revolut now runs Z.ai's GLM models on 56 Nvidia B200 GPUs and reports flat, predictable billing.
The company says coding time improved 60% and lead time 70%, with the biggest gains coming from AI handling debugging and Jira and Confluence tickets rather than writing code. There are trade-offs: GLM is text-only, so Revolut routes image and UI work to other models such as Gemini.
In short: A big fintech found it could get most of the value of premium AI coding tools for about a tenth of the cost by running an open model on its own hardware.
What it means for your business: You probably don't need 56 GPUs, but the lesson scales down: measure what your AI subscriptions actually deliver, set per-person budgets, and remember that cheaper models are often good enough for routine work.
My take: The most useful line here is that the biggest wins came from tickets and debugging, not code generation. That matches what I see: AI pays off fastest on the tedious work around the work. Self-hosting is a big-company move, but cost discipline isn't.
Source: The Register: Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted
CrowdStrike ships an "off button" for rogue AI agents
CrowdStrike announced an AI Kill Switch on Tuesday at its Fal.Con Europe conference. Part of its Falcon Guardian platform, which monitors AI agents running inside a company, it lets security teams terminate a misbehaving agent, stop it from relaunching, and cut its network access. CEO George Kurtz told CNBC, "Everyone is building AI agents, but they never had the off button, and now they do," and said it is available immediately.
Kurtz also said enterprise customers insisted on CrowdStrike protection being integrated before they would deploy Anthropic's Claude Cowork product.
In short: CrowdStrike now sells a tool that can stop an AI agent in its tracks and keep it from coming back.
What it means for your business: Whatever tools you use, every AI agent you deploy should have a clear off switch, a named owner, and limited access. If you can't answer "how would we stop it?", it isn't ready.
My take: "Kill switch" is a marketing-friendly name for something that should be table stakes. Most small businesses won't buy Falcon Guardian, but you can get 80% of the benefit by giving agents their own limited accounts that you can disable in one click.
Source: CNBC: CrowdStrike's new kill switch can stop rogue AI agents in their tracks, CEO says
Anthropic publishes FlashProve, a way to prove which AI model actually answered you
Anthropic published research and open-source code for FlashProve, a system it calls "provable inference" that lets anyone verify which AI model produced a given answer. Labs would publish a cryptographic "fingerprint" for each model that can later be matched to its outputs, without revealing the model itself, and verification is light enough to run on a local device. The idea is to stop a provider from quietly swapping in a cheaper model to save money, or a pricier one to inflate test results.
Anthropic is clear this is a work in progress, describing it as "intermediate progress, and not a completed research agenda." The spec and code are on GitHub, and there are no plans yet to deploy it at Anthropic or any other lab.
In short: Anthropic released an early, open-source method for proving that an AI answer came from the model you were promised.
What it means for your business: Nothing to install today, but if you pay for premium AI models through resellers or third-party apps, this is the kind of verification you'll eventually want written into contracts.
My take: As more apps resell AI under the hood, "which model am I actually paying for?" is a fair question with no good answer today. This is research, not a product, but it points at a real gap. I'd love to see other labs adopt it.
Source: TechCrunch: Anthropic is working on a system to prove AI models' identity
Missed yesterday's roundup? Catch up on the November 3 AI and IT news recap.