WhatsApp Help Pages Hid Secret Instructions Directing AI Bot to Avoid GDPR References
Researchers at Redact have uncovered hidden instructions in the source code of WhatsApp help pages that directed the company's AI-powered support chatbot to avoid recommending users contact data protection regulators or reference key privacy laws.
The instructions appeared in the publicly accessible code of the help section. While ordinary users saw only standard support text, the chatbot received additional internal guidance specifying topics it should not discuss. Redact examined 545 help pages and found 102 concealed blocks on 80 of them.
One directive instructed the bot not to advise users who had lost access to their phone number and could not delete their account to contact supervisory authorities. This guidance contradicted WhatsApp's own European privacy policy, which explicitly allows users to file complaints with regulators.
Additional instructions prohibited the chatbot from discussing GDPR, LGPD, or CCPA once a user mentioned legal action or an attorney, even if the user directly asked about these regulations.
WhatsApp, owned by Meta, confirmed to WIRED that the instructions were meant exclusively for its internal support chatbot. The company attributed the exposure of the blocks to a technical error and stated that the restrictions were intended to prevent the bot from providing legal advice. All hidden blocks were removed from the public code by 18 October.
Related articles
Russia Freezes Entire Hosting Networks to Block VPN Traffic, Investigation Finds
An independent test of 435 endpoints across 154 hosting providers revealed that Russian internet providers are throttling connections to entire autonomous systems rather than individual IP addresses. Networks belonging to Hetzner, OVH, DigitalOcean, M247 and HostPapa are frozen after roughly 25 TCP packets, causing downloads to stall at around 16 KB while UDP-based WireGuard tunnels remain unaffected. The blocking is applied uniformly by multiple Russian operators including Beeline, MTS and home broadband providers, indicating a centrally distributed rule from Roskomnadzor via the TSPU filtering system. Smaller providers that rent IP space from the affected autonomous systems are also impacted, regardless of their own brand or location. Cloudflare-hosted resources currently pass the test, but the list of blocked networks has remained stable since mid-2025. The researcher published an open-source script that allows anyone to reproduce the measurements in under a minute.
macOS Builds Undocumented Personal Knowledge Graph Since Version 13, Inferring Family Ties via On-Device ML
Since macOS 13 in 2022, two daemons named knowledgeconstructiond and intelligenceplatformd have been constructing a comprehensive personal knowledge graph on every Mac. The system aggregates data from Contacts, Calendar, Photos, Messages, Mail, Family Sharing, Health emergency contacts, Screen Time, and behavioral signals to map users, their relationships, locations, and activities. Results are stored in ~/Library/IntelligencePlatform as SQLite databases, including graph.db containing 181 entity classes and 278 predicates. Two CoreML models, EntityTagging_Family.mlmodelc and EntityTagging_FamilyAndFriends.mlmodelc, use 144 behavioral and photographic features to classify family and friend relationships with daily updates in entityTagging.db. An embedded LLM and embedding models further label unnamed faces and infer relationships from message text when Apple Intelligence is enabled. The graph remains queryable locally and is designed to feed personalized AI features while operating entirely on-device under Full Disk Access protections.
Breaking the Ice: Claude AI Maps TSPU Deep Packet Inspection and Censorship Bypass Methods in Russia
An analysis conducted by Claude AI examines the current state of Russia's TSPU censorship infrastructure as of November 2026, based on real-world testing from the Rostelecom network. The report details how DPI equipment blocks traffic by IP addresses, SNI inspection, protocol signatures, and active probing, with specific freezes occurring after approximately 16 KB on many foreign hosting providers. Multiple circumvention techniques are evaluated, including VLESS combined with Reality and XTLS Vision, cascaded connections through Russian entry nodes, and local tools such as zapret and GoodbyeDPI. The study highlights that large providers including Hetzner, OVH, Vultr, Scaleway, and AWS are heavily restricted, while alternatives like Leaseweb and certain academic networks remain functional. Recommendations emphasize using self-hosted domains with valid certificates, Reality-based masking, and regular IP testing to maintain connectivity. The article stresses that all information is presented strictly for research purposes.
Apple to Scan iCloud Photos and Files Shared via Public Links Starting with iOS 27.3
Apple announced that it will begin scanning content stored in iCloud when users share files or photos through public links. The feature activates with the release of iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3. Any file made accessible to anyone with the link will now be checked by Apple for malicious materials. This change effectively means that sharing via public links grants Apple visibility into the shared content for security purposes. The move aims to detect harmful files before they reach other users through shared links.