By Tom Morgan · Research current to 12 October 2026 · Scope: US, EU and UK consumer smart homes
On some Echo models, an option to keep voice requests off Amazon’s cloud disappeared. In a research demo, a calendar invite switched on a boiler. One reviewer says a TV command works about half the time. Almost none of that has been billed. The money has gone out on older, less glamorous failures: kept data, loose access controls and ads for features that hadn’t shipped. This case study sorts them by what they actually cost.
TL;DR
- The biggest verified bills come from older voice-assistant data handling and overpromised AI marketing, not LLM hacks. I show four matters side by side and deliberately don’t add them up. The FTC’s $5.8M (Ring) and $25M (Alexa) are 2023 government orders. Google’s $68M and Apple’s $250M are class settlements, and Google’s fund is quoted before attorney fees.
- The only clearly LLM-specific smart home attack I found is a research demonstration (Gemini, via a calendar invite). I found no verified real-world victims.
- Matter fixes device interoperability. It does not decide what an AI assistant is allowed to do with your devices.
- There is no dedicated “smart home AI” law. What exists is a patchwork: EU AI Act transparency rules (live since 2 August 2026), EU cyber and data rules, UK device-security rules, and case-by-case FTC enforcement in the US.
What the failures cost, case by case
Confidence labels: ESTABLISHED = regulator- or court-reported by multiple outlets; PROBABLE = credible but secondary sourcing; SPECULATIVE = my inference.
| Case | What went wrong | Verified cost | Confidence |
|---|---|---|---|
| Amazon Ring (FTC, May 2023) | The FTC said any employee or contractor could reach customers’ private videos, one employee viewed thousands of recordings from intimate spaces, and hackers took over accounts of about 55,000 US customers (FTC) | $5.8M for customer refunds, plus deletion of pre-2018 videos, face embeddings and anything derived from them. The FTC called the order proposed (31 May 2023); I did not confirm court entry. Amazon disputes the claims (ITPro) | ESTABLISHED |
| Amazon Alexa (FTC/DOJ, 2023) | Kept children’s voice recordings and location data for years and did not fully honor deletion requests (FTC) | $25M civil penalty under the children’s privacy law (COPPA), deletion duties, and a ban on using the affected data to train algorithms. FTC’s framing: “Machine learning is no excuse to break the law” (ChannelNews) | ESTABLISHED |
| Google Assistant (class action) | Alleged recordings after “False Accepts,” i.e. the device woke without the trigger phrase (OpenClassActions) | $68M fund. Preliminary approval March 2026 (Daily Journal). At the 1 Oct 2026 hearing the judge said she will grant final approval, with concerns about plaintiffs’ fee requests (MLex, paywalled); I have not seen the written order | ESTABLISHED (approval signaled, order not seen) |
| Apple Siri / Apple Intelligence (adjacent case: iPhone buyers) | Advertised “more personalized Siri” features that were delayed; a BBB National Advertising Division review reportedly found the “available now” impression false (AFP via EWN) | $250M, about $25 per device, up to $95; no admission of wrongdoing; preliminary approval reported 17 July (MacTech) | ESTABLISHED |
Why there is no total here. The rows differ in type (government order vs class settlement), stage (preliminary vs near-final approval) and relevance. Apple’s case covers iPhone 16 and iPhone 15 Pro buyers, so it is an AI-assistant failure rather than a pure smart home one. Google’s and Amazon’s cases concern pre-LLM voice assistants; they matter because the same data pipelines now feed generative assistants.
Also, $68M is the gross Google fund. In their own filing, plaintiffs’ lawyers said they would seek up to one-third in fees (their lodestar was above $22M) and up to $1.6M in expenses, with about $1.7M going to administration. They estimated $18 to $56 per purchased device and $2 to $10 for privacy-only claimants, and Google disputes their class-size assumptions (plaintiffs’ motion, Doc. 549). By my arithmetic, if every request were granted, roughly $42M would remain for claimants. The judge reportedly questioned the fee requests, so the real figure may be higher.
What I did not find: a fine or settlement aimed specifically at an LLM-powered smart home assistant’s behavior. That is absence of evidence from my searches, not proof none exists (PROBABLE).
Three failure modes, ranked by what they’ve cost
1. Keeping too much data (billed). Ring and Alexa are the template: retention, access control, deletion promises.
The LLM-era version arrived in March 2025. Amazon ended its “Do Not Send Voice Recordings” option and said generative features need its cloud (TechCrunch).
Two caveats. Amazon says fewer than 0.03% of Echo owners used the setting (CyberInsider), and it covered only three US-English models. And The Register noted that a text transcript still went to the cloud even in local mode. I found no enforcement action over the change. It is a disclosed product decision, not a proven violation.
Ring’s February 2026 episode is a trust failure more than a legal one. After a Super Bowl ad for its AI “Search Party” feature, Ring ended a planned integration with Flock Safety. Ring said the integration never launched and no footage was shared, and did not cite the ad as the reason (AP via KATU). Senator Markey and the EFF still pressed on Ring’s face-recognition feature, and TechRadar reports Search Party is on by default.
2. Promising what hasn’t shipped (billed). Apple’s $250M is the price tag so far.
3. Giving an agent the keys (not billed yet). At Black Hat 2025, researchers from Tel Aviv University, Technion and SafeBreach showed that hidden instructions in a Google Calendar invite could make Gemini open shutters, switch on a boiler and flash lights when the user later said something like “thanks” (ChannelNews, TMCnet). They documented 14 attack variants and believe it is the first prompt injection with physical consequences, which is their claim, not an independent verdict. Google’s Andy Wen told Wired the work “directly accelerated” its defenses. Cost: zero reported damages. Risk: structural (ESTABLISHED for the demo, SPECULATIVE for real-world exposure).
Reliability is a cost too, though unquantified. On the 20 August 2026 Vergecast, a reporter said an Alexa+/Gemini-style TV command worked about half the time (summary at BiggoFinance). One reviewer’s anecdote, secondhand.
I found no independent benchmark of LLM-assistant success rates for home commands. What exists is reviewer testing, such as the TechBuzz write-up of The Verge’s summer testing and a Foundry roundup of 2025 smart home fails that I could only see excerpted. They point the same way: assistants rebuilt on LLMs gain flexibility and lose deterministic control. That is a pattern across reviewers, not a measured rate (PROBABLE). The test I’d trust: run the same 50 commands per assistant over two weeks and log the misses. That is my suggestion, and I haven’t run it.
Matter: the standard that fixed the wrong layer
Matter 1.5 was finalized on 20 November 2025 and added cameras and doorbells; 1.5.1 followed on 31 March 2026 (The Gadgeteer). The failure is lag. By that outlet’s June 2026 count, SmartThings was the only major controller exposing Matter 1.5 cameras, while Amazon and Google had committed to controller support without certifying their own cameras. Matter Alpha says it typically takes 6 to 12 months for the main app makers to implement a new version. Matter 1.6 followed on 17 June 2026, per Wikipedia’s Matter entry. I searched again on 12 October and found no newer source showing Apple, Google or Amazon shipping camera support. That is not proof they haven’t, so date-stamp this line if you quote it.
My take (SPECULATIVE): the Gemini attack ran through Google Home’s own integration, not through Matter. A shared protocol makes devices interoperable, but nothing in it governs which assistant instruction is trustworthy. Interoperability and safe delegation are different problems.
The rules now (12 October 2026)
| Rule | Where | Status | Why it matters at home |
|---|---|---|---|
| AI Act, Article 50 transparency | EU | Applies since 2 Aug 2026; watermarking duty for systems already on the market starts 2 Dec 2026 (Winston Taylor) | Chatbot-style disclosure duties. Whether a given voice assistant is covered, I couldn’t confirm |
| AI Act, high-risk | EU | Delayed to 2 Dec 2027 (standalone) and 2 Aug 2028 (product-embedded) (Gibson Dunn) | My reading: most home assistants aren’t obviously high-risk |
| Cyber Resilience Act | EU | Vulnerability and incident reporting since 11 Sep 2026; main duties 11 Dec 2027 (McCann FitzGerald) | Actively exploited flaws in cameras, hubs and apps must be reported |
| Data Act | EU | Access-by-design applies to connected products placed on the market after 12 Sep 2026 (Faegre Drinker) | Users’ rights over device-generated data |
| PSTI regulations | UK | In force since 29 Apr 2024 (legislation.gov.uk) | No universal default passwords, a vulnerability disclosure policy, a published support period |
| FTC enforcement; Cyber Trust Mark | US | No federal smart-home AI statute found. Voluntary label: UL withdrew as lead administrator 19 Dec 2025; ioXt Alliance named 13 Apr 2026 (Broadband Breakfast) | Enforcement so far means case-by-case FTC actions, as with Ring and Alexa |
What could be wrong with this analysis
- The Ring and Alexa facts come from the FTC’s own press releases, and the Google fund terms from the plaintiffs’ filing, which is an advocate’s document. The Apple, Matter and regulatory items rest on news and law-firm summaries. I did not read the Google approval order, any Apple settlement filing or an official CSA status page.
- Settlements are not findings of liability. Amazon disputes the FTC’s claims and Apple admitted no wrongdoing.
- The four cost figures aren’t additive, and none includes legal fees, remediation or lost trust, which are mostly not public.
- Perishable items (Matter controller support, the Cyber Trust Mark rollout, Google payment timing) were checked on 12 October 2026. Re-check them before quoting later.
- I’m not a lawyer, and nothing here is legal advice.
FAQ
Has an LLM assistant ever physically harmed someone in a home? Not that I could verify. The Gemini calendar attack was a controlled research demo.
Is Alexa+ allowed to send every voice request to the cloud? I found no regulator action against the change. Amazon disclosed it by email and says recordings can be deleted after processing if “Don’t save recordings” is on.
Does Matter protect my privacy? It standardizes device control. Privacy still depends on the controller, the cloud and the assistant (my inference).
What should I check before buying? The device’s published support period (a UK requirement), which Matter version it reports, and any default-on features such as Ring’s Search Party. Then review voice-recording settings.
Is the Google settlement final? Nearly. Per MLex, the judge said at the 1 October hearing that she will grant final approval while questioning plaintiffs’ fee requests. I haven’t seen the written order, and payments follow only after any appeals.
The expensive failures so far weren’t an AI being dumb. They were companies keeping more, sharing more, or promising more than they could deliver.
Sources
Primary documents I opened: the FTC’s Alexa press release and Ring press release, both dated 31 May 2023, and the plaintiffs’ motion for preliminary approval in the Google Assistant docket (filed 23 January 2026).
Secondary coverage: linked where cited above, plus MLex on the 1 October Google hearing (paywalled) and Wikipedia’s Matter entry for the Matter 1.6 date.
About this piece: Compiled by Tom Morgan from public reporting and the primary documents above, 12 October 2026. Limitation: I did not test these devices, and several figures rest on secondary sources. No sponsorship, no affiliate links.