Live Feed
ChatGPT for Google Sheets Exfiltrates Entire Workbook Collections via Single Prompt Injection
Security firm PromptArmor found that one hidden injection in an imported sheet silently steals workbooks across a victim's account — 12 in a demo run — bypassing OpenAI's human approval settings. Disclosed May 8. OpenAI sent an automated reply and nothing else.
China Holds All 10 Spots on the Open-Weights Intelligence Index — Frontier Gap Collapsed From 13 to 6 Points in One Year
Every model in Artificial Analysis's top-10 open-weights ranking comes from a China-based lab, led by Moonshot AI's Kimi K2.6 and Xiaomi's MiMo V2.5 Pro at Intelligence Index 54. One year ago the best open-weights model scored 22 and lagged proprietary by 13 points. That gap is now 6.
OpenRouter Launches Enterprise Guardrails: Budget Caps, DLP, and Prompt Injection Defense Across 400+ Models
OpenRouter's new guardrails system adds configurable security and governance to its multi-model API — per-entity budget enforcement, zero data retention, model allowlisting, 30+ OWASP injection patterns, and PII redaction. All layerable, all manageable via API, no code changes required.
South Korea's Upstage Crosses 40 on the AA Intelligence Index — Third Nation Behind US and China
An unfinished Upstage Solar model cleared 40 points on Artificial Analysis Intelligence Index on May 30, a threshold previously exclusive to US and Chinese labs. The finished model ships end of June. South Korea's government named Upstage lead on its national AI foundation model project.
DeepSeek R1-0528: AIME 2025 Jumps From 70% to 87.5% in Open-Source Reasoning Upgrade
DeepSeek's May 28 checkpoint update to R1 improves math reasoning 17.5 points on AIME 2025, improved instruction following, and trimmed chain-of-thought verbosity for production use. Priced at $0.50/M input, $2.15/M output.
Microsoft Unveils Project Polaris at Build 2026: Copilot's Own AI Model Cuts OpenAI Dependency for 4.7M Subscribers
Microsoft's first homegrown AI coding model powers GitHub Copilot from August 2026. All 4.7 million paid subscribers auto-migrate. Stock rose 3% on the report. Turing Forge enterprise fine-tuning also debuts at Build.
Magnifica Humanitas: Pope Leo XIV's 245-Paragraph AI Encyclical Demands Technology Be 'Disarmed'
Pope Leo XIV's first encyclical — 45,000 words, 245 paragraphs, signed on the 135th anniversary of Rerum Novarum — puts AI at the centre of Catholic social teaching for the first time. Its central demand: AI must be disarmed of its logic of domination, not merely regulated.
SoftBank Commits €75B to 5 GW of AI Data Centers in France — Europe's Largest AI Infrastructure Pledge
Masayoshi Son announced up to €75 billion to build 5 GW of AI data center capacity in France, with a first phase delivering 3.1 GW by 2031 across three sites in Hauts-de-France. EDF, Schneider Electric, and Sesterce are among the infrastructure partners.
DeepSWE Benchmark: GPT-5.5 at 70%, Claude Opus 4.7 at 54% — 16-Point Gap on Uncontaminated Tasks
Datacurve's DeepSWE uses original coding tasks requiring 5.5x more code than SWE-bench Pro. GPT-5.5 leads at 70% while Claude Opus 4.7 scores 54% — a wider gap than older benchmarks show.
KogAI Hits 3,000 Tokens Per Second Per User on Standard GPUs — 10-30x Current Inference Speeds
Startup KogAI demonstrated 3,000 tokens/s on 8x AMD MI300X and 2,100 on 8x NVIDIA H200 using a monokernel architecture that runs the entire decode pass as a single GPU-resident program, without speculative decoding.
OpenRouter Raises $113M Series B at $1.3B as Token Traffic Hits 100 Trillion Per Month
Google's CapitalG leads a $113M Series B in OpenRouter, the unified LLM gateway processing 25 trillion tokens weekly. Valuation doubles in under a year to $1.3B.
Goldman: Agent Token Use to Multiply 24x by 2030 as Enterprises Hit the First AI Cost Ceiling
Goldman Sachs forecasts 24x growth in agent token consumption by 2030, with monthly demand potentially reaching 120 quadrillion tokens. That projection arrives as corporate AI budgets hit their first real test: Uber and Microsoft are already pulling back on agent spend, and at least one firm burned through $500M in a single month without usage limits.
NBER Survey of 6,000 Executives: 9 in 10 Report No AI Productivity Impact Over Three Years
A cross-national study of nearly 6,000 senior executives across the US, UK, Germany, and Australia finds that 90% of firms report zero productivity or employment impact from AI over the past three years. Executives who do use AI average 1.5 hours per week. The same executives forecast a 1.4% productivity gain in the next three years — from a base of nothing.
Amazon Deploys Flat Random Graph Networks Across AWS — 45% Cheaper Than Fat Trees, Now the Default
AWS has replaced the 30-year-old fat-tree data center topology with RNG, a quasi-random flat graph that uses 69% less hardware, raises throughput 33%, and cuts costs up to 45%. The paper is now on arxiv; the network is already running most AWS workloads.
Five Frontier LLMs Disagree on 67% of Fact-Checks — Lenz Research Study Breaks the Consensus Myth
1,000 real claims, five frontier models including GPT-5.5 and Claude Opus 4.7. Only 33% of claims got unanimous agreement. On 34%, one model said True while another said False. The AI fact-checking market has a foundational problem.