Original title: Shifting from Injection to Interaction: Rethinking Web Security in the Age of LLMs and Beyond
Nivedita Singh, Alsharif Abuadbba, Yansong Gao, Surya Nepal +1 · September 2026
Websites and browser tools increasingly rely on large language models, or LLMsAI systems that understand and generate human-like text for web applications.. However, adding these AI systems creates new digital dangers and makes old security flaws worse. Researchers studied how classical web weaknesses interact with new AI risks. When websites run on LLMs, old attacks like cross-site scripting become more dangerous. At the same time, AI problems like prompt injectionAn attack where hackers manipulate AI inputs to cause unintended actions.—where tricky text misleads the model—can spread through connected web software. The authors looked across client-side, server-side, and pipeline layers to evaluate existing defenses. They identified three unsolved issues: sneaky text instructions, rogue autonomous agents, and tracking threats after launch. To fix this, they designed a framework that checks inputs, protects prompts, isolates outputs, governs agents, and monitors runtime.
Imagine you live in a house with smart locks and cameras that make it safer. Now, suppose the smart devices themselves can be hacked to let intruders in, and the old broken windows still work too. In this paper, LLMs are like those smart devices, adding new vulnerabilities while old ones persist, requiring a complete security overhaul.
Web applications now use AI, creating new security risks for developers and users. Protecting online systems requires understanding how AI changes traditional threats.
The abstract does not specify empirical results or experimental validation; it is a survey that identifies unresolved challenges and proposes a framework without solving all issues.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Shifting from Injection to Interaction: Rethinking Web Security in the Age of LLMs and Beyond — Nivedita Singh, Alsharif Abuadbba, Yansong Gao et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.