In a September 28, 2026 post, GitHub Security Lab reported that its AI-guided audit taskflows had found and reported 24 vulnerabilities in Android applications. The examples concern third-party apps, including OsmAnd and the Wikipedia Android app.
Two apps, two reported risks
GitHub Security Lab described three vulnerabilities in OsmAnd and detailed one involving its exported MapActivity, which accepts data used to import settings. GitHub said another installed app—even one without permissions—could use the flaw to change settings covertly and send an attacker coordinates for loaded map tiles and route endpoints.
The Wikipedia Android app example involved a different path: a crafted deep link could direct the app’s WebView, an in-app browser component, to an attacker-controlled page. GitHub said the page could run JavaScript in the WebView and, when combined with a separate flaw in cookie-domain checks, expose long-lived cookies and enable account takeover across Wikimedia projects.
| Android app | Reported flaw | Potential impact reported by GitHub Security Lab |
| OsmAnd | An exported MapActivity accepts data used in settings import. | Another app could covertly change settings and send map-tile coordinates and route endpoints to an attacker. |
| Wikipedia Android app | A deep-link handling flaw could open an attacker-controlled page in the app’s WebView; a separate cookie-domain flaw could expose cookies. | The combined flaws could expose long-lived cookies and enable account takeover across Wikimedia projects. |
How the Android-focused taskflows guide an audit
GitHub’s Android-specific workflow separates mobile from non-mobile entry points. Another taskflow checks selected vulnerability classes in the context of each entry point and component, helping focus the model’s analysis on how an app handles those paths.
The GitHub Security Lab Taskflow Agent repository describes a multi-agent, YAML-driven framework that uses MCP. It lists Python 3.10 or Docker as requirements and identifies the project’s license as MIT. GitHub says an audit of a medium-sized repository can take one to two hours; the described workflow requires a GitHub Copilot license and uses premium model requests, which can consume substantial tokens.
Why AI findings still need security review
GitHub cautioned that model-generated findings can be false positives, rely on implausible states, or misjudge severity. It recommends that a researcher experienced in mobile security review each finding.
