Dominic Jainy is a leading figure in the evolution of automated application security, specializing in the intersection of large language models and mobile vulnerability research. As an IT professional with deep roots in artificial intelligence and blockchain, he has spent the last few years refining how developers identify critical flaws before they reach the public. His recent work highlights a shift in the industry toward “taskflow” automation—a method that moves beyond simple AI prompts to create structured, multi-stage audits of complex codebases. This interview explores his insights into the recent discovery of two dozen Android vulnerabilities, the technical mechanics of deep link hijacking, and the ongoing necessity of human intuition in an increasingly automated world. We discuss the transition from broad security scans to targeted mobile entry point analysis and how logic flaws in apps with millions of downloads can lead to silent, devastating privacy breaches.
How do you structure AI-assisted security audits to move beyond general prompts?
Instead of feeding a large language model an entire codebase and asking it to find bugs with a general prompt, we break the audit into surgical, Android-specific stages. We utilize frameworks like the GitHub Security Lab Taskflow Agent to first map out every mobile entry point—things like exported activities, services, broadcast receivers, and deep links—which creates a roadmap for the AI. Once the attack surface is defined, a second taskflow evaluates each point against specific vulnerability classes, such as unsafe broadcasts, cross-app scripting, or WebView risks. This granular approach ensures the AI understands the relevant attack surface even in repositories that contain a messy mix of mobile, web, and desktop code. By isolating these components, the model can focus its “attention” on the intricate logic of the Android ecosystem rather than getting lost in thousands of lines of irrelevant backend scripts.
What are the technical implications of the vulnerabilities found in highly popular apps like OsmAnd?
In an application like OsmAnd, which has more than 10 million downloads, the discovery that the exported MapActivity could be manipulated is a sobering reminder of the “confused deputy” problem. Researchers found that a malicious, unprivileged app on a user’s device could supply attacker-controlled values such as silent_import and replace to the app’s settings. This allows an attacker to silently modify the application’s map-tile source so that every map request is sent to a server under the attacker’s control. The terrifying part is the sensory silence of the attack: the victim keeps using the map normally, while their real-time coordinates, movements, and route destinations are being siphoned off and logged. It turns a trusted navigation tool into a covert tracking device without ever alerting the user or requiring special permissions.
Can you walk us through the complexity of the account takeover chain discovered in the Wikipedia application?
The Wikipedia vulnerability is a classic example of how a minor coding oversight, like using an endsWith() check for hostname validation, can lead to a catastrophic failure. By registering a wikipedia:// deep link that accepted any domain ending in “wikipedia.org,” the app was tricked into loading content from a malicious site like evil-wikipedia.org. This initial breach of trust was then compounded by a second domain-suffix validation flaw in the application’s cookie-handling code. An attacker could craft a webpage with a specific deep link and, once the victim taps it, the WebView provides sensitive Wikimedia cookies to the attacker’s site. This chain allows an attacker to steal usernames, long-lived authentication tokens, and session tokens valid across Wikipedia, Wikidata, and Wikimedia Commons, effectively hijacking the entire account.
Why do you believe expert human review remains non-negotiable despite these AI breakthroughs?
While AI is incredibly efficient at spotting recurring code patterns and identifying sensitive APIs, it still struggles with the nuanced logic of an application, often misjudging severity or overlooking mitigating factors. We found that forcing the model to generate a proof of concept is a great way to improve the triage process, but it is not a silver bullet. Large language models can effectively identify the “what” of a vulnerability, but they often struggle with the “how” when it comes to complex, multi-stage exploits. There is a specific kind of intuition required to understand how a bug fits into a larger threat landscape, and false positives remain a significant hurdle that only a human can clear. In the end, these tools are force multipliers for experts, not replacements for the seasoned eye of a security researcher.
What are the operational requirements for running these types of automated security agents at scale?
Transitioning to these AI-assisted workflows requires a serious commitment of resources, starting with a GitHub Copilot license to access the necessary models. For a medium-sized repository, a thorough audit can take anywhere from one to two hours and involves a constant stream of tool calls that consume a high volume of premium requests. All the findings and the traces of the AI’s logic are systematically logged into an SQLite audit_results table for later review by a human analyst. It’s a resource-intensive process that demands both high-compute power and a steady hand to manage the resulting data. You aren’t just paying for the AI’s time; you are investing in a structured pipeline that converts raw code into a prioritized list of actionable security insights.
What is your forecast for the evolution of AI-driven vulnerability research?
Looking ahead from 2026 to 2028, I expect AI agents to shift from being mere assistants to becoming autonomous “threat hunters” that operate in real-time during every stage of the development lifecycle. We will likely see these taskflows becoming even more specialized, moving beyond simple logic flaws into automated patch generation and hardware-level security verification. However, as these tools become more accessible to defenders, the barrier to entry for attackers also drops, leading to a relentless, automated arms race. The ultimate winners in this space will be the organizations that can most effectively blend the raw speed of machine learning with the creative, adversarial mindset of a human hacker. We are moving toward a future where “secure by design” is not just a slogan, but a state maintained by thousands of specialized AI agents working in the background of every commit.