OpenAI Contractors Manually Review Real User Chats in Project Lily
OpenAI has recruited hundreds of third-party contractors to manually review real user dialogues with ChatGPT. The work forms part of project Lily and gives reviewers direct access to prompts and responses that may include personal, medical, or other confidential details.
Reviewers are tasked with briefly summarizing each user query, comparing four model-generated answers, and scoring them on a scale from one to seven. They also flag issues such as excessive sycophancy, attempts by the model to appear human, and inappropriate use of emojis.
Although account names are masked and conversations pass through a Privacy Filter, OpenAI confirmed to 404 Media that the filter does not always remove all personal data. Reviewers may additionally see blocks of user memories containing prior conversation context and approximate location information.
Contractors are hired through external firms including Crossing Hurdles and Mercor, with some specialists earning more than $50 per hour. A similar human-evaluation process is used by Anthropic for conversations with Claude.
Under OpenAI policy, chats from personal plans may be used to improve models. Users can disable this by navigating to Settings → Data Controls and turning off the “Improve the model for everyone” option. The setting applies account-wide but does not retroactively remove previously processed data. Temporary chats are excluded from training, though they may still be reviewed for abuse detection.
Related articles
UDP Proxies and QUIC Protocol: How Real IP Addresses Leak Through Anti-Detect Browsers
Anti-detect browser users relying on UDP-capable proxies face a hidden risk of real IP leakage when the browser fails to properly route UDP traffic. The QUIC protocol, which powers HTTP/3, runs over UDP and enables features like 0-RTT handshakes, independent streams, and connection migration that can bypass proxy routes. WebRTC connections using ICE, STUN, and TURN further increase exposure because they often attempt direct UDP paths outside the configured SOCKS5 proxy. Without deep network stack control such as TUN interfaces or socket interception, browsers may send WebRTC and QUIC packets through the host's real network interface. Aurorium Browser claims to solve this by natively supporting UDP proxying so that both QUIC and WebRTC traffic stays inside the tunnel. The article stresses that simply disabling WebRTC or forcing HTTP/2 fallback is insufficient and can itself create detectable anomalies for anti-fraud systems.
Google to Offer Granular Controls for Advanced Protection Mode in Android 16
Google is preparing more flexible settings for its Advanced Protection security mode that first appeared in Android 16. The changes were discovered by Android Authority researchers while examining Google Play Services version 26.36.30. A new Expert features section will let users enable individual protections such as USB Protection, intrusion detection logging, and restrictions on unsafe Wi-Fi networks without activating the entire strict mode. USB Protection blocks new USB connections while the screen is locked to prevent physical attacks, though it can interfere with fast charging on Pixel 6 and newer devices. Users will also be able to opt out of automatic connections to open or risky Wi-Fi networks if they regularly use public hotspots. Intrusion Logging remains optional and stores encrypted security logs in the cloud. The update aims to preserve core security benefits while removing the all-or-nothing requirement of the current Advanced Protection implementation.
rkn-block-checker 0.6.0 Adds Local Web UI and Reduces False Positives on Anti-Bot Responses
The open-source tool rkn-block-checker received a major update in version 0.6.0 that improves accuracy when detecting Russian ISP censorship blocks. The previous version incorrectly flagged sites protected by anti-bot systems as blocked by Roskomnadzor because it searched for common stub-page strings without checking HTTP status codes. A new logic layer now separates rate-limit responses such as HTTP 429 from genuine provider stub pages that typically return 200 OK or 451. The release also introduces a fully local Web UI built with Python’s standard-library HTTP server and vanilla JavaScript, eliminating the need for heavy frameworks. Results stream in real time via NDJSON so users can watch checks complete without polling or WebSocket connections. The project remains available on PyPI and GitHub under the MIT license.
Digital Surveillance in Your Pocket: Audit Uncovers Tracking and Weak Cryptography in 11 Top Russian Android Apps
An independent security researcher conducted a static analysis and passive logcat review of 11 popular Russian Android applications from banks, government services, marketplaces, and navigation tools. The audit identified 237 vulnerabilities in total, including six apps signed with outdated RSA-1024 or DSA-1024 keys that violate current NIST standards. VK led the risk ranking with 11 trackers and a 98/100 surveillance score, while Gosuslugi transmitted analytics signals to VK, Yandex, Google, and Huawei ecosystems. Five apps used WebRTC ICE candidate leaks to detect VPN usage, potentially bypassing user privacy protections. Only Goskluch scored low at 42.4/100 by limiting trackers to two, avoiding VPN detection, and using a compliant RSA-2048 signing key. The methodology relied solely on MobSF, jadx, and logcat without bypassing SSL pinning or requiring root access.