GitHub AI agent finds 24 CVEs on Android, including GPS tracking failure

The Taskflow Agent, an open-source framework maintained by GitHub Security Lab, found 24 vulnerabilities (CVEs) in the Android ecosystem — among them failures that allow GPS tracking and content manipulation of Wikipedia. The tool does not trigger a single scan: it breaks audit in incremental steps guided by AI, which, according to the laboratory, leads to deeper results than conventional automated scans. The case was reported by Tech Times.

What did Taskflow Agent find on Android?

The balance sheet released is 24 CVEs in the Android environment. Two fault classes draw attention to the nature of the impacts: one related to GPS tracking, which exposes the user's location, and another linked to Wikipedia's content manipulation — that is, a failure capable of compromising the integrity of information displayed or served by ecosystem applications.

As with CVEs, each finding received its own identifier and follows the normal flow of responsible disclosure, with correction in charge of the maintainers of the affected projects.

How does the Taskflow Agent incremental step approach work?

The central difference of the Taskflow is in the orchestration. Instead of asking a model to "find bugs" at once, the framework decomposes the goal in chained subtasks: each step produces inputs that feed the following, allowing the agent to pursue exploration hypotheses rather than just recognizing suspicious patterns.

It is this decomposition that explains why the audit has gone deeper. Logic failures, interlinking between components and authorization problems rarely appear in a single grep — appear when someone (or something) follows the flow from beginning to end.

Why do 24 CVEs on Android matter beyond GitHub?

The result is added to a recent crop of autonomous security agents. Microsoft revealed that its autonomous AI found 16 new vulnerabilities on Windows, including 4 critical CERs, while Anthropic reported more than 2,000 unknown failures in seven weeks with Mythos.

What distinguishes Taskflow is the format: being open-source, the framework can be run by any security team, not only by those who have a contract with a large cloud provider. This pushes AI-assisted audit into the territory of small projects and individual maintainers — exactly the profile that usually stays out of corporate bug bounty programs.

There's a counterweight, though. Self-employed agents have also become a surface of attack: ESET has already warned about malicious AI agents planted in open-source repositories. Tools that read and execute code deserve the same scrutiny as the code they analyze.

What do CVEs on Android change for those developing in Brazil?

The country is among the largest Android markets in the world, and most of the apps consumed here are born from small teams, which do not maintain a dedicated security team. An open-source framework that extends the scope of the audit has a direct effect on this profile.

The regulatory cut reinforces the urgency. Location data are treated as personal data by LGPD, which transforms a GPS tracking failure into legal and not just technical risk — with a duty to respond to incidents and possible exposure of responsibility to the controller.

What's not clear about the Taskflow Agent and the CVEs yet?

The coverage does not detail the complete list of hit packages, which versions have been corrected or patch deadlines — this depends on the maintainers of each project. There are also no public figures on the agent's false positive rate, decisive metric to determine whether the scale approach outside the laboratory.

What is already established is the method: decompose, chain and deepen. The rest is to follow the CVEs to correction.

Frequently Asked Questions about Taskflow Agent and CVEs on Android

What is Taskflow Agent?

It is an open-source GitHub Security Lab framework that conducts security audits with AI agents, dividing complex goals into incremental and chained steps. It was used to identify 24 CVEs in the Android ecosystem.

What types of failures were found on Android?

Among the 24 CVEs are a GPS tracking vulnerability, which exposes the user's location, and another related to Wikipedia content manipulation. Each failure received its own identifier and follows the flow of responsible disclosure.

Why does this approach find more flaws than a common scan?

Because the agent pursues the flow of the application step by step, rather than just looking for known patterns. This allows to achieve logic and chaining failures between components, which traditional scans usually miss.