Umar Iqbal

We're less than a month out from the MADWeb 2026 submission deadline! We look forward to seeing your exciting work!

We invite both full and work-in-progress papers. Consider submitting and help spread the word!

📅 Deadline: Dec 11, 2025 (AoE)
📜 Submit: madweb26.hotcrp.com

There are a lot more details to our approach, including several open problems. If you’re interested in learning more, we encourage you to read the paper (arxiv.org/pdf/2403.04960)! You can also catch Yuhao's talk about IsolateGPT at NDSS next week in San Diego!

GitHub - llm-platform-security/SecGPT: An Execution Isolation Architecture for LLM-Based Agentic Systems An Execution Isolation Architecture for LLM-Based Agentic Systems - llm-platform-security/SecGPT

We are excited to announce that IsolateGPT will appear at NDSS 25. To foster follow-up research, we release our code (github.com/llm-platform...). Also as @llamaindex.bsky.social Pack (llamahub.ai/l/llama-pack...). Lead by Yuhao Wu w/ @franziroesner.bsky.social @yoshikohno.bsky.social & Ning Zhang

Our evaluation demonstrates that it is indeed feasible to isolate execution in LLM agentic computing paradigm: it mitigated security and privacy issues without loss of functionality and its performance overhead is under 30% for 2/3rd of tested queries

IsolateGPT runs individual tools in isolated containers, to ensure that tools cannot interact with components outside of their execution environments. Then to enable interaction between sandboxed tools, it allows apps to exchange messages only via a central trustworthy module

Our security architecture, named IsolateGPT, tackles these challenges. IsolateGPT assumes an LLM-based digital assistant that supports third-party tools for tasks, such as online shopping, email management, etc. and aims to secure adversarial manipulations between tools

Access control & isolation have existed in prior systems but their application to LLM computing paradigm is non-trivial: isolated environments need to be securely provided access to broader system context, & secure interfaces need to be defined for natural language interactions

To that end, our research has focused on adapting systems security principles in improving the security of LLM integrations and LLM-based agentic systems. We recently explored the feasibility of access control and isolation in improving the security of LLM interfacing with tools

The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions Today's LLMs are susceptible to prompt injections, jailbreaks, and other attacks that allow adversaries to overwrite a model's original instructions with their own malicious prompts.

While there is a serious emphasis on making LLMs robust, e.g., training LLMs to prioritize privileged instructions openai.com/index/the-in..., we believe that tried and-tested systems security principles have not been given similar attention in securing LLM integrations

These issues can manifest in conventional computing systems where LLMs are getting deeply integrated, such as OSs and mobile apps, and also in agentic systems (or AI agents) which interface with various tools and content on the internet

Fundamentally the key issue is that LLMs load instructions from various sources (system, user, tools) in a shared context window where without safeguards, LLMs treat them with the same privileges

At a high level, the interfacing between system components in determined at runtime based on instructions from system components – which can be untrustworthy, malicious, or compromised – such as tools developed by third-party services or arbitrary content hosted on the internet

While this execution paradigm has the potential to fundamentally transform computing, there are serious security, privacy, and safety risks!

For example if a user prompts their LLM-based personal assistant to download and store email attachments in a cloud drive, the LLM can predict the necessary interfacing between email and cloud drive tools to carry out the task

LLMs have enabled a new computing paradigm, where the system relies on ML models to resolve user queries expressed in natural language! In this paradigm, new features can be implemented via natural language specs, without requiring explicit implementation from software developers

Privacy | 2024 | The Web Almanac by HTTP Archive Privacy chapter of the 2024 Web Almanac covers the adoption and impact of online tracking, privacy preference signals, and browser initiatives for a privacy-friendlier web.

The Privacy chapter was written by a cornucopia of experts: Yash Vekaria, Benjamin Standaert, @maxostapenko.com , Abdul Haddi Amjad,
Yana Dimova, Shaoor Munir, Chris Böttger, and Umar Iqbal

almanac.httparchive.org/en/2024/priv...

Catch up on for the latest on a very important topic for the web!