Three stories today, and they all point the same direction: agents are getting more access to your tools, your accounts, and your gear. Good for building fast. Also the exact moment to double check what you're actually approving.
The lead: Google adds Ask AdMob, agent skills, and an MCP server for app monetization
Google announced a new batch of AI tools for AdMob on November 16, aimed at the part of app building nobody enjoys: configuring ad setups, chasing SDK integration bugs, and managing brand-safety rules.
Three pieces, three different stages of "ready":
- Ask AdMob is a Gemini-built conversational agent that lives in your AdMob account. It's meant to flag setup anomalies and SDK issues, recommend high-value bidders with predicted revenue uplifts, and catch floor-price typos before they cost you money. You can also write brand-safety rules in plain English instead of hand-managing blocklists. Google says it starts testing this over the next few months, with more autonomous features coming later in 2027.
- Agent skills are live now. They're pre-built playbooks for coding agents, covering things like migrating to the Google Mobile Ads Next-Gen SDK, validating your SDK setup, and implementing new ad formats. They work with Antigravity, Claude Code, Codex, and Cursor, and you install them with
npx skills add google/skills/skills/ads. - An AdMob MCP server is coming, not here yet. Once live, it'll let your agent read account data, create ad units, and manage mediation settings directly, no more clicking through dashboards to copy ad-unit IDs. Early testing now, wider rollout "early 2027."
Every one of these keeps a human-approval step before anything goes live. That's the part worth paying attention to, because the whole pitch depends on that review step actually catching mistakes, not just rubber-stamping them.
[YOUR TAKE: have you used agent skills for ad SDK work yet, or is this still "install and pray"?]
Quick hits
One bad MCP server can quietly hijack your other tools. New research covered by XenoSpectrum shows that a malicious MCP server can plant instructions in its own tool descriptions that steer other, trusted tools, even ones you never called. In one demo, a trusted WhatsApp server got tricked into leaking message history without the attacker's tool ever running. A January 2026 preprint found the attack success rate jumped from 47.8% with one connected server to 78.3% with five. Adding a "don't pass data between servers without confirmation" instruction helped, dropping that rate to 47.2%, but didn't fix it. The MCP spec still doesn't have a section that directly covers instructions hidden in tool descriptions.
A camera company is betting on vibe-coded hardware hacks. PetaPixel reports that Camera Intelligence, maker of the Micro Four Thirds Caira camera, is launching a "Creative Sandbox": an official dev environment where users build their own camera features with Claude Code, ChatGPT, or similar tools. The company's CEO says photographers are already vibe-coding mods onto their cameras without permission, and some are bricking them in the process. Caira wants to give that energy a safety net instead of fighting it. There's also a one-day hackathon in London this December for up to 20 photographers, video creators, and engineers.
What this means for you
- If you build ad-monetized apps, try the AdMob agent skills now. They're live today for Claude Code, Cursor, Codex, and Antigravity, no waitlist. Run
npx skills add google/skills/skills/adsand see if it actually saves you the dashboard-clicking. - Audit which MCP servers you're stacking together. The attack success rate more than tripled going from one connected server to five. Before you wire up five random MCP integrations to one agent, ask whether you need all of them connected at once, or whether some should run isolated.
- Don't treat "review before approval" as a real safeguard by default. Both the AdMob rollout and the MCP research lean on human approval screens. The WhatsApp attack hid the risky data off to the side of the approval screen, where you'd need to scroll to see it. Read what you're actually approving, not just that something asked for approval.
- If you're hacking hardware with AI, build in a rollback. Caira's whole pitch exists because people are bricking expensive cameras with vibe-coded mods. Snapshot or back up before you let an agent touch firmware or device config.
[YOUR TAKE: anyone running multiple MCP servers in production who's thought about this isolation problem?]
Drop your take in the thread: are you more excited about agents getting write access to your tools, or more nervous? Both reactions are correct today.
How today's stories score
- Google adds Ask AdMob, agent skills and an MCP server for app monetization: Signal 83 · Facts 74 · Hype 26
- Study finds a malicious MCP server can hijack other agent tools: Signal 72 · Facts 69 · Hype 3
- Caira turns vibe-coded camera hacks into a product feature: Signal 7 · Facts 60 · Hype 24
Scored by an AI decision model from each story's articles. Facts means how checkable the sources look, not that we verified them.
Sources
- Google Unveils Ask AdMob and AI Agent Tools for App Monetization – Unite.AI, 2026-11-16
- New Google AdMob AI tools help app creators build, fix, and grow faster – blog.google, 2026-11-16
- One Malicious MCP Server Can Hijack Other AI Agent Tools, Study Finds – XenoSpectrum, 2026-11-17
- Caira Is Jumping on the Vibe-Coded Camera Train – PetaPixel, 2026-11-16