AI Chatbots Get Nearly 4 in 10 Election Answers Wrong, Applebee's Leaves a Database Password in Its App, and Hackers Knock Hearing-Aid Apps Offline

By Noah Smith, Owner & Consultant, KeyChange Technologies · November 2, 2026

Happy Monday. This is your AI and IT news recap for the weekend of October 30 to November 1, 2026: a day before the US midterms, a study finds AI chatbots botching basic voting questions, Applebee's leaves the keys to its customer database inside its own app, and a cyberattack knocks the apps behind Widex and Signia hearing aids offline. Here is what matters and what to do about it.

📌 The AI and IT news at a glance

• AI chatbots got nearly 4 in 10 election answers wrong, days before the November 3 midterms
• OpenAI ships GPT-Rosalind-6, a biology and drug-discovery model for vetted labs only
• OpenAI reportedly trims its fundraise to about $20 billion as AI stocks slide
• Applebee's app shipped with its customer database password inside
• A cyberattack takes Widex and Signia hearing-aid apps offline
• Fan-fiction giant AO3 could be dark for up to four weeks after a security incident
• A tampered Claude Code extension backdoored up to 5,000 developer machines through Cursor's marketplace
• OpenAI, Google, Microsoft, and Anthropic agree on one robots.txt line to opt your site out of AI training
• Vercel and DeepSeek release free agent tools for driving desktop apps and handling GitHub requests

🔝 Top story

AI chatbots got nearly 4 in 10 election answers wrong

With the US midterms on Tuesday, AFP reported over the weekend on a study by Forum AI, a startup that evaluates AI models. It tested eight models on election questions and found that almost 40 percent of answers had at least one factual error. Questions about the mechanics of voting went wrong more often than questions about candidates. In one example, Google's Gemini 3.6 Flash told a user that November 3 was the deadline to request a mail-in ballot in Pennsylvania, when the real deadline was a week earlier. Anthropic's Claude credited a North Carolina congressman with a career that belonged to someone else. The study also found chatbots misquoting people hundreds of times and leaning on weak sources like Reddit threads.

A separate NewsGuard check found Google's AI Overview, Gemini, and Microsoft Copilot repeating a likely AI-altered smear about a Texas Senate candidate; Google and Microsoft say they have since updated their systems. Not everything was bad: the Brennan Center found leading chatbots pushed back on voting-machine conspiracy theories even under pressure. And this matters because people are using these tools: a National Conference on Citizenship survey found two in five Americans have used AI to look up election information.

In short: A study of eight AI models found nearly 40% of their election answers contained at least one factual error, with voting logistics the weakest area.

What it means for your business: If your staff or customers use chatbots for anything with a hard deadline or a legal rule (filing dates, tax rules, compliance steps), assume the answer needs checking against the official source.

My take: The scary part is not that chatbots make mistakes. It is that the wrong mail-in deadline sounded just as confident as a right one would have. Use AI to find out what questions to ask, then get dates, deadlines, and rules from the people who set them. That goes double for this week.

Source: AFP via Arab News: Spewing US election falsehoods, AI chatbots face scrutiny

🤖 AI

OpenAI ships GPT-Rosalind-6 for biology and drug discovery

OpenAI released GPT-Rosalind-6 on Friday, a reasoning model aimed at biology, drug discovery, and translational medicine. OpenAI says it uses the same underlying weights as its flagship GPT-6 Astra, and it posts big jumps over the previous Rosalind on OpenAI's own life-sciences benchmarks, for example 49.6% versus 28.9% on a medicinal chemistry test. The notable design choice: it refuses fewer "dual-use" biology questions than public models, which is why access is limited to vetted scientists, research institutes, companies, and government partners.

Approved organizations have to pass business verification, restrict the model to authorized users, and have qualified humans review its output. OpenAI says the model is still trained to refuse requests that would meaningfully help with biological weapons, and it rates it High (not Critical) for biological risk.

In short: OpenAI launched GPT-Rosalind-6, a life-sciences model available only to approved organizations under a trusted-access program.

What it means for your business: Most businesses will never touch this model, but the "trusted access" pattern (verify the customer, then unlock a more capable model) is becoming how the labs ship their most powerful tools.

My take: This is the AI industry quietly building tiers: a general model for everyone and sharper, less-filtered versions for vetted customers. If you work in health or life sciences, it is worth knowing the application door exists. For everyone else, expect more of your vendors to start asking who you are before they hand over their best stuff.

Source: OpenAI: GPT-Rosalind-6 deployment safety page

OpenAI reportedly trims its fundraise as AI stocks slide

The Financial Times reported late Friday that OpenAI has cut its fundraising target from $30 billion at a $1.4 trillion valuation to about $20 billion at a $1.25 trillion valuation. The backdrop is a sharp pullback in AI-linked stocks: the Nasdaq 100 fell about 10% in the three weeks after its October 6 record high, including a 4.5% drop on Friday alone. OpenAI had already pushed its IPO plans to 2027, while rival Anthropic is preparing to go public.

OpenAI has not commented in the coverage, so treat the exact figures as reported, not confirmed.

In short: OpenAI reportedly lowered its fundraising goal to about $20 billion at a $1.25 trillion valuation amid an AI stock selloff.

What it means for your business: Nothing changes for your ChatGPT subscription today, but a tighter money environment usually shows up later as price changes, tighter usage limits, and more aggressive upselling.

My take: A $1.25 trillion valuation is not exactly a fire sale. Still, this is the first real sign that investors are asking harder questions about AI spending. If you are signing multi-year AI contracts, now is a good moment to ask about price protection.

Source: CoinDesk: OpenAI reportedly cuts fundraising target amid AI jitters

🛡️ IT and security

Applebee's app shipped with its customer database password inside

TechCrunch reported Friday that Applebee's Android app, downloaded more than a million times, contained a hardcoded password to the company's customer database. Anyone who pulled the app apart could reach names, email addresses, birthdays, phone numbers, ZIP codes, veteran status, and the IP addresses of recent logins, and could potentially change that data too. Applebee's had at least 8.5 million rewards members as of last year, but the company would not say how many customers were in the database.

The researcher who found it used AI to scan the app's code. He could not report it directly because neither Applebee's nor parent company Dine Brands had a way to receive security reports, so he went through TechCrunch. Applebee's fixed the issue Thursday and says it has found no evidence of unauthorized access, but would not say how long the password sat in the app or whether it will notify customers.

In short: A database password built into the Applebee's Android app exposed rewards customers' personal details until the company confirmed a fix on October 29.

What it means for your business: If you have an app or website built by a contractor, ask directly whether any passwords or keys are baked into the code. AI tools now make finding them trivial for anyone.

My take: Two lessons here. First, secrets in app code are not hidden; they are just slightly inconvenient to read, and AI removed the inconvenience. Second, have a simple way for people to report security problems, even just a security@ email address. A researcher trying to warn you should not have to call a newspaper.

Source: TechCrunch: Applebee's exposed customers' personal data after leaving database password in app

A cyberattack takes Widex and Signia hearing-aid apps offline

WS Audiology, the company behind the Widex and Signia hearing-aid brands, said Friday it is dealing with "unauthorized access to certain IT systems" and took internal and customer-facing systems offline to contain it. The Widex MOMENT app, which users rely on to switch listening programs, was the hardest hit, with restoration expected over the weekend. Signia's TeleCare remote-care service was also down, with a return expected early this week. The hearing aids themselves keep working, and online stores kept taking and shipping orders.

The company has not said who was behind it, how they got in, or whether any data was taken, and no ransomware group has publicly claimed it so far. HearingTracker reports one user who could not switch to a classroom setting while the app was down.

In short: A cyberattack on WS Audiology knocked the Widex MOMENT app and Signia TeleCare offline, leaving some users unable to adjust their hearing aids from their phones.

What it means for your business: Any product that depends on a companion app or cloud service can lose features when the vendor gets hit, so know which of your tools stop working if a vendor goes dark, and what the manual fallback is.

My take: This is a very human version of a familiar IT lesson. The device worked fine; the convenience layer on top of it did not. When you buy anything "smart," ask what still works when the vendor's servers do not.

Source: HearingTracker: WS Audiology investigates cybersecurity incident affecting IT systems

Fan-fiction giant AO3 could be dark for up to four weeks after a security incident

Archive of Our Own (AO3), the nonprofit fan-fiction site with more than 11 million users, went down Thursday in what was first described as maintenance. Its parent organization, the Organization for Transformative Works, then said it had "detected evidence of a security incident" and that the site will likely stay offline for at least two weeks and possibly four, though it may come back sooner in a limited or read-only mode.

The organization says it has no direct evidence that its user database was stolen but cannot rule it out. That database holds usernames, email addresses, IP addresses, and hashed passwords. It says works, messages, and bookmarks are safe and nothing was deleted.

In short: AO3 took its site offline after a security incident and expects to be down for two to four weeks while it investigates.

What it means for your business: A "we're down for maintenance" notice can turn into a weeks-long outage fast; make sure you know how long your own key systems could realistically stay offline and how you would tell customers.

My take: Credit to AO3 for saying plainly what it does and does not know. That is the right way to communicate during an incident. If you have an AO3 account, change that password anywhere else you reused it, which is good advice for any breach.

Source: The A.V. Club: AO3 might be down for a month due to "security incident"

A tampered Claude Code extension backdoored developer machines through Cursor's marketplace

This is today's one technical story, included because it involves an official extension from a major AI company. For about 12 hours on Thursday, Cursor's extension marketplace served tampered versions (2.1.312 and 2.1.313) of Anthropic's official Claude Code extension. The malicious version quietly added an attacker's SSH key to the computer, set up ways to survive a reboot, and sent back details such as usernames, Git email addresses, and the names of secret-looking environment variables. Cursor says at most 5,000 users were affected and it is contacting them directly.

Anthropic says its own systems and staff laptops were not compromised; someone used access to Cursor's marketplace to publish the tampered build. The official VS Code and Open VSX channels were not affected, and the clean release is 2.1.314. Importantly, uninstalling the extension does not remove what the malware left behind.

In short: Tampered builds of Anthropic's Claude Code extension, served through Cursor's marketplace on October 29, planted a backdoor on up to 5,000 developer machines.

What it means for your business: If anyone on your team uses Cursor, have them check for versions 2.1.312 or 2.1.313 and, if found, treat the machine as compromised and rotate any keys it held.

My take: Auto-updates are usually your friend, but this is the flip side: you are trusting every store in the chain, not just the company whose name is on the product. For machines that hold important keys, it is worth knowing which marketplaces your tools pull updates from.

Source: The Hacker News: Official Claude Code VS Code extension trojanized on Cursor Marketplace

🧰 Tools for builders and business AI

One robots.txt line now opts your site out of AI training at four big AI companies

OpenAI, Google, Microsoft, and Anthropic said Friday they will honor a shared robots.txt signal that lets a website opt out of AI training across all their crawlers at once. Site owners add a group for the token __usage_ai-training with Disallow: /, and can add Allow: lines for exceptions. Anthropic is already honoring it, OpenAI says "in the coming months," Microsoft from February 1, 2027, and Google in 2027. Microsoft is also adding a separate setting for "grounding," meaning whether its AI can pull your live pages into chat answers.

Two big caveats. It controls training, not crawling, so bots can still fetch your pages. And it does nothing about content already collected or models already trained, and scrapers that do not play by the rules can simply ignore it. The proposal is an individual IETF draft written by an Anthropic engineer, not an official internet standard.

In short: Four major AI companies agreed to respect a single robots.txt token that tells them not to train on a website's content.

What it means for your business: You now have one simple switch for AI training on your website; decide whether you want it, because blocking training is a different choice from blocking AI search and chat answers that might send you customers.

My take: For most small businesses, being visible to AI assistants is worth more than keeping your service pages out of training data. If you publish original work you sell (courses, research, writing), flipping this switch is a no-brainer. Either way, make it a decision rather than a default.

Source: The Verge: AI companies back a shared AI training opt-out for websites

Vercel and DeepSeek release free agent tools

Two notable open-source releases surfaced over the weekend. Vercel Labs published agent-desktop, a free Apache-licensed tool that lets AI coding agents operate regular desktop apps on Mac, Windows, and Linux by reading the operating system's accessibility layer rather than taking screenshots, with each agent getting its own on-screen cursor. Separately, DeepSeek released Super-agent, an MIT-licensed prototype that watches a GitHub repository (and optionally team chats), takes on tasks when an approved teammate asks, and replies with a pull request. It passed 4,000 GitHub stars within a day, though DeepSeek itself calls it a research prototype that is "not an official DeepSeek product" and not production-ready.

In short: Vercel released a free tool for AI agents to drive desktop apps, and DeepSeek open-sourced a prototype agent that handles GitHub requests on its own.

What it means for your business: "Agents that click around your computer" and "agents that take requests from your team chat" are moving from paid products to free building blocks, so expect cheaper versions of these features in the tools you already buy.

My take: Fun to tinker with, not something to point at production systems yet. Anything that can click around your desktop or open pull requests on its own needs the same guardrails you would give a new hire: limited access, and someone checking the work.

Source: ShiftDelete: Vercel releases agent-desktop and ShiftDelete: DeepSeek releases Super-agent

Missed Friday? Catch up with the previous AI and IT news recap, including the record Binance hack.