Gemini key leak and 3D push
Security researchers found hardcoded Google API keys inside Android apps that could expose Gemini access across apps with about 500 million installs, raising abuse and billing risks for developers. (ciso.economictimes.indiatimes.com) At the same time, Gemini is expanding its surface: Google added interactive 3D model and simulation outputs and published extensive notebooks and a cookbook for developers, widening both capability and attack surface. (tbreak.com) (lightbook.org)
A Google API key is supposed to work like a valet key: it opens one narrow door, not every room in the building. CloudSEK said 22 Android apps shipped 32 hardcoded Google keys that now reach Gemini endpoints, and those apps add up to more than 500 million installs. (cloudsek.com) The twist is that many developers treated keys starting with “AIza” as low-risk because they were already embedded in Android apps for maps, analytics, or other Google services. CloudSEK and SecurityWeek reported that once Gemini was enabled on the linked Google Cloud project, those same keys could be used against Gemini without the app developer intending it. (securityweek.com) That turns an old mobile habit into a new artificial intelligence problem. A key copied out of one app can let someone send prompts, burn through quota, and run up charges on the developer’s Google Cloud project instead of their own. (infosecurity-magazine.com) CloudSEK said the exposed access was not limited to basic text generation. Its researchers wrote that some keys could touch uploaded files, fine-tuned models, and cached prompt data tied to the same project, which is closer to finding the office filing cabinet unlocked than just borrowing the front-door key. (cloudsek.com) Google’s own documentation says hardcoding a Gemini key should only be temporary and that the safer pattern is server-side calls where the key stays confidential. The same documentation also points developers to ephemeral tokens for some client-side Live Application Programming Interface use cases, which is Google’s way of handing out short-lived wristbands instead of permanent master keys. (ai.google.dev) Google Cloud’s broader key guidance is even plainer: don’t include keys in client code, restrict what each key can call, and delete keys you do not need. Those controls matter more when one project can hold billing, models, prompts, and multiple Google services under the same roof. (cloud.google.com) At the same time, Google is making Gemini do more things in more places. Google announced that Gemini can now generate interactive 3D models and simulations that users can rotate and adjust, turning a plain answer into a small app running inside the response. (theverge.com) Google is also widening the on-ramp for builders. Its Gemini API cookbook on GitHub now serves as a large library of examples and notebooks, and the official cookbook page sends developers straight into those hands-on guides. (github.com) (ai.google.dev) Those two moves fit together in an awkward way. The more Gemini can generate code, simulations, files, and multimodal outputs, the more expensive and sensitive a stolen key becomes, because the attacker is no longer limited to asking a chatbot a few text questions. (ai.google.dev) (theverge.com) So this story is not just “someone left secrets in an app.” It is that old Android shortcuts collided with a much bigger Gemini platform in April 2026, right as Google was expanding model families, developer tooling, and interactive output formats across the same ecosystem. (ai.google.dev) (cloudsek.com)