GitHub Security Lab has demonstrated the practical value of AI in vulnerability research, reporting 24 security flaws in Android applications discovered through its open-source AI security agent. The findings, published September 28th, showcase how specialized taskflows can guide language models to find complex vulnerabilities that might otherwise be missed.
How the AI Agent Works
The GitHub Security Lab Taskflow Agent allows security researchers to automate and package effective AI prompts into reusable workflows. Rather than relying on generic model capabilities, the taskflows break research into incremental steps, helping AI systems identify nuanced security issues faster.
Key components include:
- gather_mobile_entry_point_info.yaml: Identifies attack surfaces in mobile applications by distinguishing mobile entry points from other code paths
- classify_application_local.yaml: Guides the LLM to check for mobile-specific vulnerability classes like confused deputy attacks, insecure broadcasts, and intent-based flaws
The taskflows combine strict prompting with repeated runs to catch obvious vulnerabilities while allowing AI creativity to surface complex issues.
Critical Findings
Among the 24 disclosed vulnerabilities, two stood out:
OsmAnd Navigation App: A critical flaw allows malicious apps to track user location without consent. The app exported a MapActivity that could be triggered with specially crafted intent extras, enabling silent settings imports that could exfiltrate location data. The app has over 10 million downloads.
Wikipedia Android App: Additional vulnerabilities were found in the Wikipedia Android application, though details are still being disclosed as vendors patch.
Accessibility
The taskflows are open-source and available for other security researchers. Running them requires a GitHub Copilot license, as the prompts use premium model requests. The team estimates runtime of one to two hours for medium-sized repositories.
GitHub Security Lab has published the Android-specific audit taskflows, enabling developers to run similar security scans on their own applications. The approach represents a shift toward AI-assisted security research, where human expertise guides automated analysis rather than replacing it.