GitHub Security Lab's open source AI agent uncovers 24 Android vulnerabilities

GitHub Security Lab has detailed how its open source Taskflow Agent, built as a way for security researchers to automate and share effective AI prompts and workflows, was used to audit Android applications. In a blog post, the lab's Kevin Stubbings said he had reported more than 20 vulnerabilities in Android apps using the taskflows, which the lab says can find vulnerabilities that LLMs would have missed entirely or taken longer to identify.

The taskflows are open source and available in the seclab-taskflows repository. Running them requires a GitHub Copilot license and uses premium model requests, and Stubbings warned that the workflows can result in many tool calls that consume a large amount of tokens. Researchers can start a codespace, wait for it to initialize, and run a mobile audit script against a repository. On a medium-sized repository the process can take an hour or two, after which it opens an SQLite viewer with results in an "audit_results" table.

Because Android apps have their own specific classes of vulnerabilities, the taskflows were adapted for them. A new taskflow, gather_mobile_entry_point_info.yaml, separates attacker-controlled entry points into mobile and non-mobile categories, allowing the AI to work on repositories containing different application types while still understanding the correct attack surface. A second taskflow, classify_application_local.yaml, lists popular vulnerability classes and asks the model to consider them in the context of each entry point. For intent-based entry points, for example, it checks for common issues such as confused deputy attacks and insecure broadcasts.

The lab maintains an advisories page where these vulnerabilities are disclosed as they are found.

Analysis

Why This Matters

Lowers the barrier to automated security auditing: the taskflows are open source and can be run by any researcher or developer with a Copilot license.
Shows a practical use of AI in security: guided prompts can surface Android-specific flaw classes, such as confused deputy attacks and insecure broadcasts, that generic scans might overlook.
Twenty-four reported Android vulnerabilities means downstream fixes and patches for affected apps, which could matter for users of those applications.

Background

GitHub Security Lab is a team focused on finding vulnerabilities in open source software and improving supply chain security. Large language models have become increasingly capable of reading and analyzing code, but they are non-deterministic and can miss issues unless guided. The Taskflow Agent is an open source framework that packages AI prompts and workflows so researchers can share effective approaches. This work applies that approach to mobile security, where vulnerability classes differ from traditional web or desktop software.

Key Perspectives

Security researchers: The open source taskflows offer a reusable, shareable way to automate parts of Android auditing and to document effective prompts. Android developers: Newly disclosed vulnerabilities in the lab's advisories create pressure to review and patch affected code, and the tooling may be used by others to find bugs in their own projects. Critics and skeptics: Running the taskflows is token-intensive and requires a commercial Copilot license, which may limit adoption. The non-deterministic nature of LLMs also means results can vary between runs, and the approach is a supplement to, not a replacement for, traditional analysis.

What to Watch

GitHub Security Lab's advisories page for new disclosures as taskflows are applied to more repositories.
Whether the taskflow approach extends to other vulnerability classes or platforms beyond Android.
Community adoption of the seclab-taskflows repository and reports from outside researchers running the tooling themselves.

ZOTPAPER