Google’s new Home Speaker makes the small smart speaker interesting again. The harder question is whether a more capable Gemini-powered control point should also get more authority over the house.

That is the practical tradeoff this morning. CNET’s hands-on test describes Google’s first new smart speaker in years as one of the strongest speakers at its size. Google’s product announcement lists a $99.99 price, 360-degree sound, multi-command voice control, a physical microphone-mute switch, and optional premium features that can search Nest camera history. Better hardware makes the speaker easier to place at the center of a room. The AI layer makes the permissions behind that placement more important.

The thesis is simple: buy the speaker for the jobs it does well, then grant its AI features only the authority those jobs require. Convenience should expand one reversible action at a time—not become blanket control over locks, cameras, alarms, and occupancy routines.

What is really changing

1. The $100 speaker has to earn its place as hardware

CNET’s review matters because it evaluates the product as a speaker, not merely as a Gemini endpoint. A small device that sounds good enough for a kitchen, bedroom, or office is more likely to remain plugged in and centrally located. That makes it useful for music, timers, announcements, and routine voice control before any advanced AI feature enters the decision.

Google’s own announcement supplies the product facts but should be read as a vendor claim, not an independent test. Google says the speaker has balanced 360-degree audio, adaptive microphone processing, and the ability to pair two units with a Google TV Streamer. CNET’s independent listening test is the better evidence for audio quality; Google’s post is the source for the $99.99 price, June 25 retail availability, feature list, and microphone-mute control.

That separation is useful for buyers. First ask whether the hardware solves a real room-level need. Then ask whether Gemini’s conversational features are worth enabling. A smart-home device should remain useful even when its most ambitious cloud feature is limited, unavailable, or deliberately turned off.

2. Gemini turns the speaker into a broader control surface

Google says Gemini for Home can interpret natural speech, accept multiple commands in one request, remember short conversational context, and handle corrections mid-sentence. The company also places camera-history search and home summaries in Google Home Premium.

Those capabilities reduce setup friction, but they also widen the control surface. “Dim the kitchen lights and set a timer” is convenient because both actions are easy to reverse. Asking a speaker about a back gate, a camera event, or a security routine touches more sensitive household state.

This is where feature lists stop being enough. Buyers and installers need to distinguish between what the assistant can understand and what it is allowed to do. Natural language is an interface. It is not, by itself, an authorization policy.

3. Promptware is a demonstrated risk, not a reason for panic

CNET’s promptware report explains the household version of indirect prompt injection: malicious instructions can be hidden in content an AI assistant reads, such as messages, invitations, or documents. If the assistant can also operate connected devices, a poisoned instruction can try to cross from information into action.

The underlying SafeBreach and Tel Aviv University research paper documents 14 attack scenarios against Gemini-powered assistants across five threat classes. The researchers reported digital and physical consequences, including manipulation of home appliances, and disclosed the findings to Google. Crucially, the paper also says Google deployed dedicated mitigations and that reassessed risk fell substantially.

That last point matters. The research supports the design lesson; it does not justify claiming that the disclosed attacks still work against today’s product. The right response is neither complacency nor alarm. It is to treat any assistant that reads outside content and controls devices as a security boundary that deserves explicit permissions and confirmation rules.

4. The useful AI home is a bounded AI home

Smart-home platforms already separate devices into rooms, homes, accounts, and routines. AI control needs a similar risk model.

Low-consequence actions—music, timers, lights, announcements, and many climate scenes—can tolerate more conversational convenience. High-consequence actions—door locks, garage doors, alarm modes, cameras, and presence-sensitive automations—deserve narrower triggers, stronger confirmation, and a non-AI fallback.

Matter can improve device interoperability, but it does not decide whether an assistant should unlock a door after interpreting ambiguous context. That decision lives in the platform, account, routine, and permission layers above the device protocol.

Builder and buyer lens

For a buyer, the clean decision order is hardware, platform fit, privacy controls, then AI authority. CNET’s audio test and Google’s $99.99 price make the new speaker plausible on ordinary product merits. Google’s microphone-mute switch is a useful physical control. Neither point answers how much access a particular household should give Gemini.

For an installer or tinkerer, inventory the action graph: which voice request can trigger which routine, and which routine can affect which device? A single broad routine can quietly turn a harmless voice endpoint into a path to security-sensitive equipment.

Keep sensitive actions explicit. If a platform offers confirmation, role restrictions, activity history, or separate household permissions, use them. Preserve a manual way to operate essential devices. And avoid designing a home where one compromised account, message stream, or conversational context can operate everything.

What to do next

1. Buy it for the room, not for an AI promise

Use CNET’s listening assessment to decide whether the speaker fits the room. Use Google’s documentation to check price, supported features, premium requirements, microphone controls, and TV pairing. Do not assume a future AI feature will rescue hardware that is poorly placed or unnecessary.

2. Audit the assistant’s highest-impact actions

Open Google Home—or the equivalent app for Alexa, HomeKit, SmartThings, or Home Assistant—and list everything reachable through voice and routines. Start with locks, garage doors, alarms, cameras, and presence automations. Remove broad access that is not delivering clear daily value.

3. Keep untrusted content away from powerful actions

The promptware research focused on assistants processing invitations, emails, and shared documents. Limit automatic processing of unfamiliar content, keep software current, and require a person to confirm sensitive actions. These steps will not eliminate every AI security risk, but they reduce the chance that untrusted text can quietly become device control.

The takeaway

Google’s Home Speaker looks like a credible $100 room speaker again, which is good news for a category that had become stale. Its Gemini features also make it a more consequential part of the home.

The winning setup is not the one that gives AI the most control. It is the one that gives a good speaker enough authority to be useful, keeps sensitive actions behind clear boundaries, and still works when the AI layer is muted or unavailable.