🕵️♂️ 8M Users’ AI Chats Secretly Sold by ‘Privacy’ Extensions
A shocking exposé of how seemingly trustworthy browser extensions violated user privacy at scale. This case underscores the urgent need for stronger app store vetting and accountability in the AI era.
A Koi investigation revealed that the Urban VPN Proxy browser extension, boasting over six million users, covertly collected AI chat data from platforms such as ChatGPT, Claude, and Gemini. The extension injected malicious scripts, intercepted API traffic, and transmitted user prompts and responses to BiScience, a data broker. Marketed as a ‘privacy’ tool and even featured by Google, it monetized this data for analytics. Related extensions from the same publisher used identical hidden code, compromising over eight million users across Chrome and Edge.
🔗 Read more 🔗
📊 A Quarter of US-Trained Scientists Eventually Leave
🔗 Read more 🔗
⚙️ TLA+ Modeling Tips for Reliable Distributed Systems
A concise yet insightful guide for engineers using TLA+ to reason about concurrency and distributed algorithms—bridging formal rigor with real-world modeling practice.
Murat Demirbas shares practical strategies for building robust TLA+ models: keep models minimal, use declarative specifications, manage visibility carefully, define atomic actions precisely, and strengthen invariants and progress properties. He advises against over-constraining specifications and promotes intentional error injection to validate invariant soundness.
🔗 Read more 🔗
🏗️ IronFleet: Formally Verified Distributed Systems at Scale
A landmark in formal verification, IronFleet demonstrates that large-scale distributed systems can be both mathematically proven correct and practically efficient—a bridge between theory and production reliability.
Microsoft Research’s IronFleet project introduces a proof-driven approach to verifying real-world distributed systems. By blending TLA-style refinement with Hoare logic, the framework guarantees safety and liveness for complex systems like Paxos-based replication and sharded key-value stores. Verified implementations—IronRSL and IronKV—achieve both provable correctness and competitive performance through Dafny-based proof automation.
🔗 Read more 🔗