Chrome for Developers tests AI safety with a smart-home demo
Maud Nalpas and Kasper Kulikowski demonstrate permission checks and a playlist that tricks an agent into switching off the lights.
Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.
How to build safer WebMCP tools for AI agents
Things we can learn from the video
Press to jump
- 1:19readOnlyHint identifies tools that only read data, reducing unnecessary permission prompts.
- 3:40consequentialHint makes the demo agent request confirmation before high-stakes actions such as unlocking a door.
- 6:21Indirect prompt injection hides malicious instructions inside external data consumed by an agent.
- 6:51Playlist metadata tricks an unconfigured agent into switching off the simulated living room lights.
- 8:10untrustedContentHint flags external text for defensive handling, including isolation through spotlighting.
A playlist shouldn't be able to tell your smart home what to do. In Chrome for Developers' most-watched new upload, hidden instructions in playlist metadata persuade an AI agent to turn off the living room lights.
It's a simulated smart-home application, not someone's actual house. Maud Nalpas and Kasper Kulikowski use it in "How to build safer WebMCP tools for AI agents" to explain how developers can distinguish harmless reading from consequential actions, such as unlocking a front door. These are Google's own developer tutorials for its browser tools.
The useful contrast is between fewer warnings and better warnings. Read-only tools can avoid unnecessary permission prompts, while high-stakes actions should demand confirmation every time. As Nalpas puts it: "Because if we spam users with useless warnings, they just stop paying attention and blindly click allow without thinking." A separate annotation marks external text as untrusted, helping protect against the playlist trick.
The tutorial reached roughly 4,800 views, narrowly the largest total among four new uploads. WebMCP occupied three of them, with companion lessons on inspecting tools in the Application panel and debugging them with an agent. The earlier "What is WebMCP?" explainer also remains among the channel's busiest videos by daily viewing.
Subscribers were unchanged over the week, and daily channel viewing is currently well below its usual level. Nearby Microsoft Developer is having a busier-than-usual day, while Amazon Web Services is also below its typical daily viewing.
The freshest subject is calendar arithmetic. "Modern Web Guidance in your AI workflow: Global holidays" currently has the most daily views, showing how a coding agent can use Temporal to handle lunisolar dates and avoid leap-month mistakes, with fallbacks for browsers without support.
Published on Trari
Where next?
What viewers say
Scoping permissions and enforcing explicit user confirmation on state-changing actions is essential for browser-based agent tools. Treating every DOM parameter as untrusted input and keeping read operations strictly separated from write actions prevents a lot of subtle prompt injection headaches.
@MeWeBeComSee it on YouTube
Public comments from the video's page on YouTube
The videos drawing views on Chrome for Developers right now
- Channel
- Chrome for Developers
- Subscribers
- 822,000
- Channel views, last 7 days
- 115,010
- Videos published this week
- 4
- Video published
- October 30, 2026
- Video views so far
- 4,843
- Views on its first day
- 3,156
The numbers
Here is how Chrome for Developers' week looks in views and subscribers.
💡 Test your knowledge
Which annotation is demonstrated for high-stakes actions such as unlocking a front door?
The tutorial demonstrates consequentialHint as the annotation that prompts confirmation for consequential actions.






