The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

SearchLeak severity comparison showing Microsoft scored CVE-2026-42824 at 6.5, NVD scored it at 7.5, and EchoLeak's Microsoft CNA score was 9.3.

SearchLeak turns Copilot’s trust boundary into the bug

SearchLeak is a one-click M365 Copilot exploit chain. It shows why agent security has to move below prompts, into render and egress controls.

Lars Cornelissen · Jun 16, 2026

Snowflake

Snowflake Hybrid Tables get faster without request fees

Snowflake Hybrid Tables are row-store tables for OLTP-style work in Snowflake. New preview optimizations report up to 8x throughput.

Lars Cornelissen · Jun 15, 2026

Snowflake

Snowflake Iceberg ADLS writes are finally usable

Snowflake Iceberg ADLS support is the GA path for Azure teams to read and write externally managed Iceberg tables without moving storage.

Lars Cornelissen · Jun 15, 2026

WorkBench agents outcomes showing GPT-4 at 43 percent correct and 26 percent harmful actions, Claude Opus 4.8 at 89 percent correct and 2.5 percent harmful, and the WorkBench repo reporting Claude Fable 5 at 92 percent correct and 1.9 percent harmful.

Research

WorkBench agents close the workplace reliability gap

WorkBench agents now solve 89 percent of workplace tasks with 2.5 percent harmful actions, changing the risk math for builders.

Lars Cornelissen · Jun 15, 2026

AgentPerf relative benchmark lifts showing NVIDIA H200 at 1x, B200 MLPerf at 3.1x, GB200 MLPerf at 3.4x, and GB300 AgentPerf at 20x.

AgentPerf puts Blackwell’s agent lead at 20x per watt

AgentPerf is a benchmark for concurrent AI agents. Its first results put NVIDIA GB300 NVL72 at up to 20x Hopper efficiency.

Lars Cornelissen · Jun 13, 2026

olmo-eval context chart showing MMLU scoring gaps from OLMES: Llama3 70B at 79.8 versus 60.7, Mistral 7B at 64.0 versus 50.3, OLMo 7B 0424 at 54.4 versus 42.4, and OLMo 7B at 28.3 versus 40.5.

Engineering

olmo-eval brings statistical discipline to LLM loops

olmo-eval is an open workbench for iterative LLM evaluation. It makes tiny checkpoint gains harder to mistake for progress.

Lars Cornelissen · Jun 13, 2026

Amazon data centers WUE compared with Amazon at 0.12 L/kWh, Microsoft at 0.27 L/kWh, and the industry average at 0.84 L/kWh.

Amazon data centers put 2.5B gallons on the AI bill

Amazon data centers used 2.5 billion gallons of water in 2025. Treat that disclosure as a roadmap risk for AI products.

Lars Cornelissen · Jun 12, 2026

Bar chart of content tactics and their effect on source visibility in AI answers: citing sources +40 percent, adding quotations +38 percent, adding statistics +37 percent, fluent writing +15 percent, keyword stuffing minus 3 percent.

Generative engine optimization: how to rank inside an LLM

Generative engine optimization (GEO) is the practice of getting your content cited inside AI answers from ChatGPT, Perplexity and Google's AI Overviews. Here is what earns a citation and what to change on your site.

Lars Cornelissen · Jun 12, 2026

Engineering

Miasma worm: live coverage of the Red Hat npm attack

Miasma is a self-propagating npm worm. It hijacked Red Hat's GitHub Actions OIDC trusted publishing to ship 96 backdoored @redhat-cloud-services versions whose preinstall hook runs a Bun credential stealer that then spreads with the secrets it steals.

Lars Cornelissen · Jun 12, 2026