Chrome Adds On-Device Gemini Nano While Ask Gemini Sends Page Content to Google Cloud
Google has added an AI on device toggle in Chrome settings that activates models running directly on the user's computer. The change supports several use cases, including detection of fraudulent websites. Several gigabytes of Gemini Nano weights now reside in the browser profile directory and execute on CPU or GPU.
Users might expect the downloaded model to power local features such as reading and summarizing the current page without cloud calls. However, Google chose a different approach. The company operates two separate systems under the single name Gemini.
Two Gemini Systems
Ask Gemini / Gemini in Chrome is a cloud service. When a user requests a summary or answers based on open tabs, the page content is sent to Google infrastructure. Gemini Nano is the local model downloaded by Chrome and invoked only through specialized browser APIs by websites, extensions, or the browser itself for initial checks.
- Ask Gemini / Gemini in Chrome – runs in Google cloud, invoked by the user
- Gemini Nano via built-in APIs – runs on the computer, invoked by sites and extensions
- Gemini Nano in Safe Browsing – performs primary analysis locally, invoked by Chrome
In the last case the model itself operates locally, yet with Enhanced Protection enabled its analysis results may be forwarded to Google Safe Browsing.
The Ask Gemini button in the interface triggers the cloud service. Page content and URLs are temporarily stored in the user's Google account, and with history saving enabled they appear in Gemini activity logs. Up to ten additional tabs shared by the user may also be processed in the cloud.
Related articles
Amnezia VPN Survives Coordinated Russian Censorship Campaign Targeting AmneziaWG Protocol Fingerprints
Amnezia VPN has published a detailed post-mortem on the multi-wave blocking campaign conducted by Russian authorities against its Amnezia Free and Amnezia Premium services during June and July. The company describes a shift from simple protocol blocking to sophisticated fingerprinting of AmneziaWG traffic combined with infrastructure DDoS attacks and automated IP-subnet blacklisting. Engineers closed multiple detection vectors including zero-length UDP packets, fixed-size keepalive messages, handshake timing patterns, and nonce zero bytes. The incident forced accelerated migration to AmneziaWG 2.0, discontinuation of legacy client support, and development of AmneziaWG 3.0 while expanding VLESS infrastructure as a backup. Self-hosted users largely avoided direct protocol blocks but still faced subnet-level restrictions. The report highlights how Roskomnadzor now applies cumulative scoring across multiple traffic features rather than single definitive markers.
Data Masking: 8 Critical Questions Businesses and Developers Ask About Protecting Sensitive Data
Garda expert Dmitry Larin addresses common challenges in data masking during a recent webinar titled 'Data Masking: Battle of Opinions'. The discussion covers why masking remains essential even when encryption is deployed, how to preserve application functionality after anonymization, and the performance trade-offs of processing large databases such as 5 TB PostgreSQL instances. Different masking types including static, dynamic, selective, and streaming are explained with specific use cases for DevOps pipelines, external contractors, and BI systems. The article also examines why machine learning alone is insufficient for discovering personal data and why custom scripts fail at scale across heterogeneous environments like PostgreSQL and Oracle. Practical recommendations include combining masking with encryption, using deterministic transformations for deduplication, and separating replication from masking tasks to avoid production impact.
MAX Desktop Client Tested for VPN Detection on Windows, No Tracking Signs Found
A Habra user named Slava_B conducted an experiment on September 8, 2026, to determine whether the MAX desktop client on Windows could detect or route traffic through a VPN configured at the router level. The setup used a Keenetic router that directed Russian resources directly while sending other connections via an OpenConnect tunnel to a European VPS, with no VPN client or virtual adapter present in Windows itself. Monitoring tools including Process Monitor, Wireshark, TCPView, and tcpdump revealed that MAX.exe and MAX-service.exe processes communicate locally and connect to MAX/ONEME infrastructure along with AppTracer services. The application repeatedly accessed MachineGuid, computer name, proxy settings, device IDs, and microphone/camera information, though these reads may support diagnostics and anti-fraud functions. No connections appeared on the VPN interface, and the client did not attempt to reach IP-checking services, Telegram, or WhatsApp. The researcher noted that TLS traffic was not decrypted, so actual transmission of identifiers could not be confirmed, and results apply only to this router-based configuration.
PII-Guard: Open-Source Detector for Personal Data in Russian Text
Andrey Ivanov, an NLP researcher at red_mad_robot, has released PII-Guard, an open-source system that detects and masks personal data in Russian text before it reaches language models. The tool combines rule-based checks with a fine-tuned ruBert-base NER model to handle names, addresses, phones, passports, INN, SNILS, bank cards and other entities. It replaces detected PII with structured XML-like tags that preserve grammatical information such as gender and entity ID, allowing models to generate coherent responses that are later restored with real values. The hybrid pipeline first applies normalization, pattern matching, Luhn and weighted checksum validation, and context windows with positive and negative keywords, then merges results with model predictions via an arbitration module. Evaluation on four public datasets, including Hivetrace, alexen2 and alrosait, shows PII-Guard outperforming other open solutions on both strict span matching and type-overlap micro-F1 metrics. The project, including datasets and code, is available on GitHub and aims to reduce leakage risks while maintaining downstream model utility.