GitHub reports 24 Android vulnerabilities found with AI-guided audits

GitHub Security Lab says it found and reported 24 vulnerabilities in Android applications using its Taskflow Agent and targeted auditing workflows. Its report published September 28, 2026 describes how structured prompts guided models toward logic flaws spanning multiple application components. The total reflects vulnerabilities reported by the team as of publication.

Android-specific prompts focus the audit

The workflows first separate mobile entry points from other application surfaces, then ask the model to consider vulnerability classes relevant to each entry point and component. For intent-based entry points, those include confused-deputy issues and insecure broadcasts.

GitHub describes combining explicit checks with broader prompts across repeated runs, rather than relying on a single open-ended audit.

Two disclosed examples

The report presents two already-disclosed examples that illustrate the importance of tracing interactions between components:

• OsmAnd: The navigation app exposed an activity called MapActivity, which handled settings files and deep links and accepted settings-related intent extras. GitHub describes a vulnerability that allowed malicious apps to track the device’s location.
• Wikipedia: GitHub describes a chain involving deep-link handling and a permissive domain-suffix check used to select cookies. In the reported attack, a victim clicked a deep link on a malicious webpage. The Android app then loaded an attacker-controlled page and sent authentication cookies, enabling account takeover, according to GitHub’s technical account.

The engineering lesson is to follow the path from externally supplied input through validation to sensitive actions or data—not just inspect methods in isolation.

Open-source workflows, premium requests

The Android taskflows are available as open source in the seclab-taskflows repository. GitHub’s documented setup starts a codespace and presents results in an SQLite viewer. It requires a GitHub Copilot license and uses premium model requests. The report warns that numerous tool calls can consume substantial numbers of tokens.

Findings still need expert triage

GitHub also reports that models returned findings requiring highly unlikely conditions and often estimated severity incorrectly. It recommends reviewing each finding with a security researcher who understands mobile applications.

For teams adopting these workflows, discovery and prioritization should remain separate steps. Treat generated findings as investigation leads: reproduce the behavior, examine mitigating factors, and assess impact before assigning release-blocking severity.