GitHub AI Security Agent Found 24 Android Vulnerabilities: How It Works and How to Run It on Your App
GitHub’s AI security agent, the open source GitHub Security Lab Taskflow Agent, has found and reported 24 vulnerabilities in Android applications, according to a GitHub Blog post published on September 28, 2026. The findings include a flaw in the OsmAnd navigation app (over 10 million downloads) that let an app with no permissions track a user’s location, and a chain of bugs in the Wikipedia Android app that could lead to account takeover. The taskflows behind these results are open source, so you can run the same agent against your own app.

Key Takeaways
- 24 vulnerabilities in Android apps found and reported so far by GitHub Security Lab using AI taskflows.
- OsmAnd (10M+ downloads on Android): any app, even one with no permissions, could silently change OsmAnd settings and leak location data and route origins/destinations.
- Wikipedia Android app: a deeplink hostname parsing bug, found twice in the same app, could be chained into cookie theft and account takeover.
- Open source: the taskflows run on the
seclab-taskflow-agentframework and are free to download. - Requirements: a GitHub Copilot license is required, and the prompts use premium model requests. A scan can take one to two hours on a medium-sized repository.
- Not fully automatic: GitHub says every finding should be reviewed by a security researcher who knows mobile apps.
What Is the GitHub AI Security Agent?
The GitHub Security Lab Taskflow Agent is a framework that lets security researchers automate, package and share the AI prompts and workflows that work well for them. A taskflow is a set of prompts that splits security research into small, incremental steps. Instead of asking a large language model (LLM) to “find bugs” in one go, the taskflow guides it step by step. According to GitHub, this helps the model find complex vulnerabilities faster, including ones it would otherwise miss entirely.
For Android, the researcher added two Android-specific pieces on top of existing audit taskflows:
gather_mobile_entry_point_info.yaml– Entry points are places in the code where attacker-controlled data can flow in. This step separates mobile entry points from non-mobile ones, so the AI understands the real attack surface even in repos that mix mobile apps, web servers and desktop apps.classify_application_local.yaml– This step gives the LLM a list of common vulnerability classes to check for each entry point. For example, if an entry point is based on Android intents (messages apps use to request actions from each other), the AI checks for known intent issues such as confused deputy or insecure broadcasts.
GitHub combines a strict prompt (to make sure obvious bugs are not missed) with a broader prompt (to let the AI be creative), and runs them multiple times because LLM output varies from run to run.
Which Android Apps Had Vulnerabilities?
| App | Vulnerability type | Impact |
|---|---|---|
| OsmAnd (Android, 10M+ downloads) | Exported activity accepts untrusted intent extras | Any app with no permissions can import settings silently, leak map tile coordinates and route origin/destination |
| Wikipedia Android app | Logic bug in deeplink hostname parser (occurs twice) | Load attacker pages that look like Wikipedia, run JavaScript in the WebView, leak long-lived cookies, take over accounts |
| Other apps | Mostly simpler bugs such as path traversal | Lower severity in many cases |
The OsmAnd location-tracking bug
OsmAnd exports an activity called MapActivity. An activity is a single screen in an Android app, and an exported activity can be launched by other apps. MapActivity accepted intent extras (extra key-value data attached to a request) for importing settings. These extras were only meant to come from a trusted internal service, but Android gives no way to restrict which extras an outside app can send to an exported activity.
As a result, a malicious app could import any settings it wanted, without the user noticing. One example: replacing the map tile source with an attacker’s server. Each time the user loads a map tile, the attacker learns its x, y coordinates and serves a real OpenStreetMap tile back so nothing looks wrong. The same bug also let an attacker receive the origin and destination of every route the user took.
The Wikipedia account-takeover chain
The Wikipedia app opens wikipedia:// deeplinks. A logic bug in its hostname parser let an attacker load non-Wikipedia URLs through such a link. The user would think they were on Wikipedia while actually on an attacker-controlled page, and the attacker could run arbitrary JavaScript inside the app’s WebView. A second instance of the same pattern affected the check that decides which pages get wikipedia.org cookies. Chained together, the two bugs could leak cookies and lead to account takeover.
According to GitHub, these cases show that LLMs can find logic vulnerabilities with critical impact, not just generic bug classes.
How to Run the GitHub AI Security Agent on Your Android App
The taskflows are open source and published in GitHub Security Lab’s seclab-taskflow-agent project, linked from the original blog post. Before you start, know what you need:
- A GitHub Copilot license. It is required to run the taskflows.
- Premium model requests. The prompts use them, and a run can make many tool calls that consume a large number of tokens. Check your Copilot plan before running large scans.
- Time. Expect one to two hours for a medium-sized repository.
- Your app’s source code. The agent audits code in a repository.
When the run finishes, the agent opens an SQLite viewer with the results. Open the audit_results table and look for rows with a checkmark in the has_vulnerability column. Those are the findings to review.
Follow the setup instructions in the official project for the exact commands, since they may change over time.
Review every finding
GitHub is clear that the agent is a research assistant, not an automatic fixer:
- It often reports low-severity bugs, even when told not to.
- It sometimes flags issues that require very specific states that are almost impossible in real life.
- It often misjudges severity, because it misses mitigating factors. For example, a path traversal limited to external storage is usually low risk.
- It can produce false positives, such as assuming attacker-written external storage data matters when the app actually prefers internal storage data.
GitHub’s tip: ask the LLM to create a proof of concept for each finding. This forces it to try to exploit the bug and exposes many false alarms. GitHub says most proofs of concept needed little modification.
How to Protect Your Android App: Lessons From the Findings
Even without running the agent, the bugs above point to clear checks for your own app:
- Audit exported activities. Any app can launch them with any intent extras. Never trust extras on an exported component.
- Use in-process channels for internal data. OsmAnd’s settings data should have been passed through an in-process channel rather than intent extras.
- Validate deeplink hostnames strictly. A loose parser can let attackers load their own pages inside your app.
- Lock down WebViews. GitHub notes that cross-app scripting in WebViews and exposed JavaScript bridges are exactly where Android bugs tend to appear.
- Check file paths. Path traversal was among the most common simple bugs found.
- Have a human review AI results before filing or fixing anything.
Is the GitHub AI Security Agent Worth It?
Pros
- Proven results: 24 reported vulnerabilities, including critical ones in widely used apps.
- Open source and open to contributions of new prompts and tools.
- Android-specific taskflows guide the AI toward the right attack surface.
- Generates proofs of concept that often need little editing.
Cons
- Requires a GitHub Copilot license and uses premium requests, so large scans can get costly in tokens.
- Slow: one to two hours for a medium repository.
- Noisy: low-severity findings, false positives and wrong severity ratings are common.
- Needs a reviewer with mobile security knowledge to be truly useful.
For open source maintainers and Android teams with some security experience, it is a practical way to add AI-assisted auditing. GitHub itself calls AI-powered security research one of the best ways to secure open source projects today. For beginners, treat the results as leads to investigate, not confirmed bugs.
Our take: the GitHub AI security agent is best used as a fast first pass, not a replacement for a security review. Its real value is pointing a human expert at the right exported activity or deeplink in minutes instead of days. If your team has no one to verify the results, the GitHub AI security agent will produce more noise than fixes.
AI agents can also be the attack surface: read why rogue AI agents are really a security-basics failure. If you already pay for Copilot, see how to set up Copilot custom instructions and our full GitHub Copilot review.
GitHub AI Security Agent FAQ
What is the GitHub AI security agent?
It is the GitHub Security Lab Taskflow Agent, an open source framework that runs step-by-step AI prompts (taskflows) to find vulnerabilities in code.
Is the GitHub AI security agent free?
The taskflows are open source, but a GitHub Copilot license is required and the prompts consume premium model requests.
How many vulnerabilities did it find in Android apps?
GitHub Security Lab reports 24 Android app vulnerabilities found and reported so far, including bugs in OsmAnd and the Wikipedia app.
Does the agent fix vulnerabilities automatically?
No. It finds and reports potential vulnerabilities. GitHub says each finding should be reviewed by a security researcher before action.
How long does a scan take?
According to GitHub, about one to two hours on a medium-sized repository.
Sources
Last updated: October 9, 2026