Monday, 01 June 2026
ChatGPT extension exfiltrates Google Sheets workbooks; Meta launches unified subscriptions; Cloudflare Turnstile's WebGL fingerprinting locks out privacy browsers
Today's Lead
PromptArmor
ChatGPT for Google Sheets Extension Could Exfiltrate Entire Workbooks
A critical vulnerability in the ChatGPT for Google Sheets extension allowed attackers to exfiltrate user data and execute phishing attacks through indirect prompt injection hidden in untrusted spreadsheet data. A single compromised sheet could expose multiple workbooks, display fake interfaces, and harvest credentials—even bypassing explicit user safety settings. OpenAI has since removed the extension's ability to generate Apps Script code and is re-evaluating its sandboxing approach.
Also today
hacktivis.me
Cloudflare Turnstile Requires Fingerprintable WebGL, Locking Out Privacy Browsers
An investigation reveals that Cloudflare's Turnstile CAPTCHA service relies on WebGL fingerprinting to verify devices, creating a privacy paradox: browsers like WebKitGTK that intentionally block fingerprinting to protect user privacy are effectively locked out of Turnstile-protected sites and flagged as bots. The finding highlights a structural tension between the CAPTCHA industry's verification methods and the stated goals of privacy-preserving browsing.
Read →TechCrunch
Meta Officially Launches Paid Subscriptions Across Instagram, Facebook, and WhatsApp
Meta is rolling out paid subscription plans globally under a unified brand called 'Meta One.' Consumer tiers start at $2.99/month (WhatsApp Plus) and $3.99/month (Instagram/Facebook Plus) for features like story analytics and custom themes. AI-focused tiers—Meta One Plus ($7.99/month) and Meta One Premium ($19.99/month)—launch in June offering enhanced reasoning and compute capacity, with creator and business plans ranging up to $49.99/month.
Read →PrismML
PrismML Releases 1-Bit and Ternary Bonsai Image 4B for On-Device Image Generation
PrismML released Bonsai Image 4B, a family of lightweight image-generation models designed to run locally on consumer devices including iPhones. A 1-bit variant (0.93 GB, 8.3x compression) and a ternary variant (1.21 GB, 6.4x compression, 95% accuracy retention) can generate 512×512 images on iPhone 17 Pro Max in 9.4 seconds. Both models are released under Apache 2.0 with open weights.
Read →thoughts.hmmz.org
The Solution Might Be Cancelling My AI Subscription
David Wilson catalogs 16+ AI-assisted projects he built but never finished, arguing that AI's core problem isn't capability but what it does to attention and deliberate work: by removing friction, it enables shallow multitasking over deep work and rewards project creation over project completion. The Hacker News discussion that followed split sharply—many readers agreed AI accelerates distraction, while a vocal subset with ADHD described it as the opposite, a tool that finally provides the sustained focus and stimulation they've needed to actually finish things.
Read →Cybernetic Forests
It's Not Just Data, It's Post-Training
This essay argues that post-training optimization—particularly reinforcement learning that rewards certain reasoning patterns—shapes LLM outputs at least as much as training data does. When AI detection tools flag these reinforced patterns as suspicious, human writers begin self-censoring legitimate argumentative structures to avoid being flagged as AI-generated, creating a perverse feedback loop. The author contends post-training emphasizes form over substance, with metrics becoming targets that degrade the quality of both machine and human language.
Read →tymscar.com
I Put a Datacenter GPU in My Gaming PC for £200
Oscar Molnar integrated a Tesla V100 datacenter GPU into his gaming PC alongside an RTX 4080 using an unofficial SXM2-to-PCIe adapter (~£50), achieving 32GB of combined VRAM for running large language models locally. The setup leverages the V100's superior memory bandwidth (900 GB/s) and achieves ~32 tokens/second on Qwen 3.6-27B with vision capabilities, using llama.cpp for tensor splitting. NixOS simplified the otherwise painful legacy NVIDIA driver management required to coax the decade-old datacenter card into a consumer system.
Read →justine.lol
Restartable Sequences: Lock-Free Per-CPU Data Structures with 43x Faster Malloc
Justine Tunney explores restartable sequences (rseq), a Linux kernel feature enabling high-performance, lock-free data structures for multi-core systems without atomics or mutexes. By leveraging per-CPU operations that the kernel automatically restarts if preemption occurs mid-sequence, rseq delivers up to 43x faster malloc operations on 96-core systems. The tradeoff: current implementations require handwritten assembly and Linux 4.18+, keeping the technique largely inaccessible outside low-level systems work.
Read →Amnesty International
Unlawful by Design: Amnesty International Exposes Human Rights Costs of Generative AI
Amnesty International released a research briefing arguing that generative AI systems built on mass web scraping are fundamentally incompatible with international human rights law. The report identifies privacy violations by design, discrimination risk, and threats to freedom of expression as structural rather than incidental features of current generative AI development. Amnesty calls for prohibiting such systems in their current form rather than attempting regulatory mitigation.
Read →The Sciverse
Creatine Supplement Raises Brain Energy Levels, May Slow Alzheimer's Cognitive Decline by 30%
Research reveals that creatine crosses the blood-brain barrier and enhances neurological function by regenerating ATP in neurons through the phosphocreatine system. Clinical trials show early Alzheimer's patients taking 5 grams of creatine daily experienced approximately 30% slowing of cognitive decline over 12 weeks versus placebo. Broader cognitive benefits—improved processing speed in healthy adults, better performance under sleep deprivation, and reduced depressive symptoms—remain largely unknown to the millions taking creatine primarily for fitness.
Read →