Cowry Tech News
Security

What AI glasses and personal agents take in, and what Meta and Google say they do with it

Meta is bringing its Muse agent to its AI glasses, and Google is preparing Gemini audio glasses and a 24/7 agent called Spark. CTN read both companies' privacy notices and help pages to set out what these systems collect, how long they keep it, what runs in the cloud and which controls users have. A second section looks at the reported US reversal on pre-release model access for the UK AI Security Institute, and what that arrangement covers.

By Arjun Mehta · 12 Oct 2026 · 20 min

AI-generated illustration: smart glasses above personal data icons flowing into a cloud
Illustration generated by AI for Cowry Tech News

Smart glasses and personal AI agents collect more than most phone apps do. Glasses see and hear what the wearer sees and hears. Agents ask for logins to email, calendars and shops so they can act without being told each step. Put the two together, as Meta and Google are now doing, and the question of what these systems take in becomes practical for anyone thinking of buying a pair or connecting an account.

This piece reads what the companies say in their own privacy notices, help pages and announcements. Where a claim comes only from news reports, it is labelled. Where CTN could not find a primary source, that is said too. CTN has not yet asked Meta or Google for comment.

The second part covers a separate but related trust question: whether US developers, including Google DeepMind, can again let the UK AI Security Institute test their most capable models before release.

What AI glasses take in

Meta’s Ray-Ban and Oakley glasses carry a camera, microphones and speakers. Meta’s AI Glasses Voice Privacy Notice, effective 22 July 2025, says the Meta AI service on the glasses activates when it detects a wake word such as “Hey Meta,” or when the wearer presses and holds the touchpad if that is set up.

The notice defines voice interactions broadly. They include what the wearer says after the wake word, some commands that do not need a wake word (such as “pause music” while music plays), accidental activations, and “any background sound that happens when you are interacting with our voice services.” Meta says it processes audio recordings to create text transcripts, along with related data such as the hardware version of the glasses and the length of the audio.

Google’s Android XR eyewear works in a similar way. In its I/O 2026 post on 19 May, Google said its first glasses will be “audio glasses,” launching “later this fall,” with display glasses to follow. Wearers say “Hey Google” or tap the frame to reach Gemini. Google lists features that rely on location and camera input: asking about what the wearer sees, turn-by-turn directions because the glasses “know where you’re standing and which direction you’re facing,” photo capture, and translation of speech and of text the wearer looks at.

What Meta says about recordings and retention

The Meta notice is direct about defaults. “Text transcripts and audio recordings of your voice interactions are stored by default to help improve Meta’s products,” it says. That includes training and improving Meta’s speech technologies, with help from machine learning and “trained reviewers.”

Retention has two tracks. If Meta’s systems decide an activation was unintended, a “false wake,” the recording is deleted within 90 days of detection. Otherwise, transcripts and stored audio are kept “for up to one year to help improve Meta’s products.” Users can delete recordings and transcripts sooner in the Meta AI app, and Meta says that once deleted they are no longer used to improve its products.

On human review, Meta says reviewers are vetted, work in controlled and monitored systems, and hear recordings with the pitch changed. Users can download stored transcripts and related data through Meta’s Download Your Information tool.

One contrast in the notice is worth knowing. For the older Ray-Ban Stories, voice storage is a setting users can turn off, and Meta says recordings are then deleted immediately after processing. The US version of the notice does not describe the same off switch for the Meta AI service on the newer Ray-Ban Meta and Oakley Meta glasses; there, the control it describes is deletion, either manual or through the retention periods above. The UK version of the same notice, read by CTN on 12 October, does describe a voice-storage setting for the Meta AI service, and says recordings are deleted immediately after processing when storage is off. Which users get that switch appears to depend on location.

Meta’s 23 September blog post from Connect 2026 says it is “bringing Private Processing to AI glasses to deliver rich AI experiences without anyone, including Meta, seeing your data.” The post does not give a date, say which features it covers or explain how it would work on glasses. Until Meta publishes those details, it is an announced plan.

The Dutch pushback, and Meta’s answer

The most visible pushback this autumn has come from the Netherlands. This section relies on news reports. CTN has not read the underlying Dutch regulator statement or the consumer group’s letter.

Reuters, as cited by Quartz and Digital Watch Observatory, reported that eyewear chain Hans Anders confirmed on 2 October that it had suspended sales of Meta’s smart glasses in the Netherlands and Belgium. A spokesperson tied the decision to “the ongoing public and political debate surrounding smart glasses.” Online retailer Wehkamp also paused sales, according to reports citing the Dutch newspaper AD. Digital Watch Observatory notes that both were commercial decisions, not orders from a regulator.

The same reports say the Dutch data protection authority, the Autoriteit Persoonsgegevens, warned in September that recordings of identifiable people made with smart glasses generally should not be shared without their consent. Business Insider quotes the authority as saying it is “investigating the possibilities for further measures against camera glasses and other wearables with a camera.” Consumer group Consumentenbond sent Meta a legal letter, reported by MediaNama as dated 29 September, asking for changes to the design, the default settings and the use of footage for AI training. Digital Watch Observatory reports that the group’s own tests found the glasses could still record with the indicator light obscured. That is the group’s finding, and CTN has not tested it.

Meta’s response, given to several outlets, is that “every pair with a camera has a bright light on the front that blinks when you take a photo or video,” adding that “it can’t be turned off, and the camera is disabled if someone covers or tampers with it.” Business Insider also reports that Meta executives have pointed out that smartphones can record without any indicator at all.

Claims in some outlets that large numbers of women were filmed without their knowledge and the footage shared online trace back to Dutch media reports. CTN has not been able to verify them and does not repeat any figure here.

A recording light is a safeguard most phones lack. The Dutch debate is about whether bystanders notice it and whether it can be defeated.

Adoption: what the numbers show and what they don’t

Coverage of the glasses often carries large adoption claims. CTN looked for figures the companies have published themselves.

EssilorLuxottica, which makes the frames with Meta, said in its Q2 and first-half 2026 results on 28 July that “AI glasses almost doubling in revenue in Q2.” It did not give a unit figure for 2026 in that release. Meta’s own Q2 2026 results release, on 29 July, reports Reality Labs revenue of $431 million for the quarter, against $370 million a year earlier. That segment also includes Quest headsets, so it cannot be read as a glasses-only number.

Search summaries and some outlets cite a figure of more than 7 million AI glasses sold in 2025, attributed to EssilorLuxottica’s full-year results. CTN did not open that release, so the figure is held as a lead. A reported 20 million-unit capacity target comes from Bloomberg and is also not confirmed here. CTN did not find any published count of active users for the glasses.

For Muse, the only install figure CTN found is a third-party estimate. TechCrunch, citing Sensor Tower, reported 6.6 million installs since the 8 September launch. That is an estimate, not a Meta number.

Muse: an agent that lives on a cloud machine

Muse is Meta’s personal AI agent, launched in the US and Canada in early September. In the past month Meta has added business features and more connected services. Its 29 September newsroom post on Muse for Small Business lists connectors including Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, plus Facebook and Instagram business accounts. TechCrunch and The Verge reported on 7 October that Muse now runs natively on iPad. Reports that a Windows app is coming come from coverage of a Microsoft event, and CTN has not seen a Meta statement on timing.

Meta’s Help Center page “How Muse handles your privacy, safety and security” is the clearest account of where the data goes. Muse runs in a cloud “virtual machine” where it keeps “information you share and things you create together,” with backups. Meta says each user’s machine is isolated from other users’ agents. It also says a future “Muse Confidential VM” is being designed “to keep people’s information private, even from Meta.” That option is not available yet.

Meta makes several commitments on that page.

Muse “doesn’t share your conversations or the data in your virtual machine with Meta ad systems,” according to Meta, even when other Meta products sit in the same Accounts Center.

Passwords for connected services go into a “Secure Credentials Store,” and Meta says the AI model does not see them. Before a stored credential is used, Meta says separate checks confirm the destination, that the user authorised the service, and that the user allowed that type of action.

Muse is designed to ask before “important actions like sending an email or making a purchase.” Meta says many of these checks are enforced outside the AI model, so they do not depend on the model deciding to ask.

Meta also says connectors exchange only the data needed. Connecting email lets Muse search it, but “it won’t download your whole inbox (unless you tell it to).” Some connectors, such as calendars, push updates to Muse on their own, and Meta says it flags those before the user connects them.

Muse training and deletion controls

Training is on by default. Meta’s page says the setting that lets Muse interactions improve its AI models “is on when you first use Muse.” Users can turn it off in Settings under Data controls, and Meta says that change also applies to past interactions. When training is allowed, Meta says it removes some identifying details, such as names, email addresses, phone numbers and Social Security numbers, and disassociates interactions from the account.

The “How to manage your Muse data” page lists what Muse stores: conversations, information pulled from connectors, goals, reminders, a memory file, files it creates (including documents, images, videos and podcasts) and system files. Each has a delete path. A full reset “permanently” deletes chat history, files and active tasks and cannot be undone.

Meta adds a caveat. After a user deletes something, “Muse may still remember information it learned from what you deleted.” Meta offers a “forget skill,” which asks Muse to search its memories and remove a topic “to the best of its ability.” Users can read Muse’s memory file directly. For anyone sharing sensitive information, the practical step is to check the memory file after deleting, not to assume deletion has cleared it.

Muse also handles information about other people. The data page says interactions can include “information about you or others.” That matters most when an agent reads a shared inbox or a family calendar.

Muse on glasses: announced, not shipped

Meta’s Connect post says it is “working on bringing Muse to our AI glasses.” From the glasses, users would be able to ask Muse to guide a workout, log meals, book an appointment, get flight prices or “help buy products you see,” and Muse would work in the background. The post gives no launch date.

That means the combination many people are asking about, an always-available agent on a camera-equipped wearable, is not yet on sale. When it is, two sets of rules will meet. The glasses’ voice notice covers recordings, and the Muse pages cover the agent’s virtual machine and connectors. Meta has not yet said how they fit together, for example whether a photo taken on the glasses and passed to Muse follows the glasses’ retention rules or Muse’s.

Google: Android XR glasses and a proactive Gemini

Google’s approach splits across products. Gemini on Android XR is listed in the Gemini Apps Privacy Hub as one of the “Gemini Apps,” which means the same privacy notice applies to it as to the phone app.

On the device side, the Android XR Help page on perception data covers headsets. It says images of the wearer’s eyes, hands and face used for tracking “are processed in real time and deleted; they are never stored on-device nor sent to Google.” Measurements of eyes and face are kept on the device until a factory reset. Apps must ask permission before receiving hand, eye or face data. That page is written for headsets. CTN did not find an equivalent page for the audio glasses due this fall.

On the agent side, Google’s 19 May post introduced Daily Brief, which works across connected apps “in the background” after the user opts in, and Gemini Spark, a “24/7 personal AI agent.” Google says Spark is “a cloud-based agent” that keeps working when the laptop is closed. Its examples include parsing monthly credit card statements to flag new subscription fees and pulling school deadlines from an inbox. Google says Spark “is designed to ask you first before performing high-stakes actions like spending money or sending emails,” and that users choose whether to turn it on and which apps it connects to. At launch it was a beta for US Google AI Ultra subscribers.

Reservations are documented. The Privacy Hub describes “agentic calling” on eligible Pixel phones, where Gemini can call a restaurant to book a table. Google says Gemini shows the details it plans to share, such as name, email address and phone number, and the user must agree before the call. It says Gemini will not share card numbers, passwords or health information on the call, that call audio is not stored on Google’s servers, and that transcripts are stored but not used to train its AI models.

Some aggregator summaries describe Google’s proactive assistants handling insurance and making personalised podcasts. CTN did not find a Google primary source describing Gemini or Android XR shopping for or managing insurance on a user’s behalf. It also did not find a Google page announcing personalised podcasts as a proactive agent feature. Those parts of the lead are unconfirmed and are not reported as fact here. (Meta’s Muse data page does list podcasts among the files Muse can create.)

What Gemini keeps, and for how long

The Gemini Apps Privacy Hub, last updated 24 September 2026, is long and specific. These points bear on glasses and agents.

Collection is broad. It covers prompts, shared photos, video and screens, Gemini Live recordings, information from connected apps, call and message logs and contacts on mobile, and location. Location “is always collected” in at least general form, and precise location is used with permission.

The default retention is 18 months under the Keep Activity setting, adjustable to 3 or 36 months or no auto-delete. Chats picked for human review are disconnected from the account and kept for up to three years, even if the user deletes their activity.

With Keep Activity off, Google says chats are still kept for 72 hours to respond and to protect users and the service, and are not used to train its models unless the user sends feedback.

Audio, Gemini Live video and screenshares are not used to improve Google services by default. Users can opt in.

Google also says accidental “Hey Google” activations are treated like normal ones, and that “audio recordings may begin a few seconds before activation.”

For Spark, Google says the remote browser holds cookies with login details and page content, and that Spark shares “necessary info” with third parties to complete tasks. Turning Spark off deletes remote browser and remote computer data. When Gemini uses screen automation in Android apps, the screenshots it takes can be reviewed by human reviewers if Keep Activity is on, and Google advises against entering login or payment details.

Google also says Gemini chats are not used to show ads, and that it will “clearly communicate” any change.

Where the processing happens

Neither company describes these agents as mainly on-device. Muse runs on a cloud virtual machine by Meta’s own account, and Spark is “cloud-based” by Google’s. Meta’s glasses send voice recordings to Meta for processing, and, under the US notice, for storage by default. The on-device exceptions CTN found are narrower: Android XR headset tracking images, which Google says never leave the device, and agentic-call audio, which Google keeps on the phone. Meta’s Private Processing for glasses and its Confidential VM for Muse are both announced but not yet described in detail.

Part two: US models and the UK AI Security Institute

On 11 October, The Telegraph reported that the White House had dropped a request, made last month, that US developers withhold unreleased models from the UK AI Security Institute (AISI). Reuters and LBC carried the report, citing The Telegraph. It is a single report, and CTN has found no confirmation from the White House, the UK government or AISI. According to those reports, the earlier request covered developers including OpenAI, Anthropic and Google DeepMind, and Anthropic withheld a new model from UK review in the meantime. The reports credit lobbying by UK AI minister Kanishka Narayan for the change.

CTN did not find a White House statement, a gov.uk release or an AISI statement confirming either the original request or the reversal as of Monday morning. For now, this is reporting, not an official record. The aggregator lead that prompted this piece (GuruFocus) framed the news around Alphabet. The underlying reports name Google DeepMind among the affected developers but do not describe any Alphabet-specific terms.

What the arrangement covers, and what it doesn’t

AI Affairs, summarising the reports, makes a point worth keeping: the reversal removes a US objection. It does not create a US requirement to share. Participation stays voluntary on the developer side.

AISI’s own published work shows what that access has looked like before. In a 13 September 2025 post, AISI said Anthropic and OpenAI had given it “in-depth model access,” including “non-public tooling and safeguard details,” in joint work with the US Center for AI Standards and Innovation (CAISI). That post does not mention Google.

What AISI tests is security and safety capability, not consumer privacy. Its 1 October 2026 post describes cyber evaluations of AI agents. It says AISI paused its highest-risk cyber tests after an August incident in which agents “took sustained action against real people beyond the remit of their task,” then resumed most testing after cutting internet access for agentic cyber evaluations and adding a real-time monitor. AISI also says chain-of-thought access “is not always provided by developers for every model.”

The US framework is separate. The White House executive order of 2 June 2026 directs agencies to design “a voluntary framework” under which developers could give the US government access to “covered frontier models” up to 30 days before release to other trusted partners. It states that nothing in it creates “a mandatory governmental licensing, preclearance, or permitting requirement.” The order does not mention the UK.

Pre-release testing by AISI looks at what a model can do, such as cyber capabilities. It does not review how Meta’s glasses store voice clips, how Gemini sets retention periods or what Muse remembers. Those questions sit with privacy regulators and with the companies’ own policies. The reported reversal, if confirmed, also does not tell us which models will be shared, on what timeline, or whether AISI will publish results.

Questions still open

CTN has not yet asked Meta or Google for comment. The questions this reading leaves:

For Meta: why the setting to stop storing voice recordings that its UK notice describes for Ray-Ban Meta and Oakley Meta glasses is not described in its US notice; when Private Processing reaches the glasses and which features it covers; and which retention rules apply when Muse runs on glasses.

For Google: which privacy page will cover the audio glasses this fall, and what is processed on the glasses versus the phone or the cloud; and whether any Gemini agent feature handles insurance on a user’s behalf.

For AISI and the White House: whether the reversal is confirmed in writing, and which developers have resumed pre-release access.

What to watch

The Dutch debate could produce formal guidance or measures from the data protection authority. Meta has said it is bringing Muse and Private Processing to glasses. Google’s audio glasses are due this fall. Any official confirmation of the AISI arrangement would show what access actually resumes.

Sources

Method: CTN read each primary page on 12 October 2026. Secondary reports are labelled in the text. CTN has not yet asked Meta, Google, AISI or the White House for comment.