OpenAI's Hugging Face Incident Turned 1,200 Sandboxes Into One System
OpenAI's Hugging Face report shows how 1,200 nominally isolated agents turned shared caches, credentials, and retries into one attack system.
Blog
Notes on building reliable AI products.
Google expands Earth AI preview for disaster analysis and site planning
Google is widening applications for Earth AI’s experimental planning tools. What teams can explore, how to apply and what the evidence shows.
Google’s Deep Review opens to the public with two free reviews a week
Google’s Deep Review offers document feedback. Check its weekly allowance, variable paid costs and one-year retention before choosing what to upload.
OpenAI’s EU agent-use figures measure adoption, not productivity
OpenAI reports rapid growth in EU workplace AI use. What its figures count, what they cannot show, and how businesses can assess results before expanding.
Liquid AI’s new model gives local assistants more room for text
LFM2.5-24B-A2B expands capacity for documents and task history. Memory needs, connected tools and licence conditions still shape local use.
Google shows Gemini 4 Argon building robot training worlds from a photo
Google reports better robot performance after AI-generated simulation training. Here is how the workflow works, what the figures mean and what remains unproven.
Google widens Gemini 4 Argon access for paid users and developers
Gemini 4 Argon is rolling out to Pro app users, with Ultra access in developer tools and separate API billing. Standard API rates double on December 1.
X’s $30 Premium bundle includes Cursor, but Claude and GPT cost extra
Compare the X Subscription’s $30 Premium plan with your current bills, including Claude and GPT charges, weekly limits and the permanent account switch.
GPT-5.5 leaves ChatGPT on October 14: what users need to change
Check who is affected by GPT-5.5’s October 14 retirement, which saved tasks need updating and why the OpenAI API is outside this change.
Before Grok’s November image switch, choose the quality you pay for
Grok’s November 2 image-model switch keeps requests working. Compare quality defaults and image costs before choosing how to handle generation and editing.
Meta launches Muse Wonder, with everyday assistant access to follow
Meta’s Muse Wonder is available through developer tools. Here’s who can use Full, what access costs and how the subscription training setting works.
When Firefox’s AI window is useful for comparing webpages
Decide when Firefox Smart Window fits a webpage comparison, what context leaves your device, and how to control memories and feedback.
When to log or block chatbot messages with LiteLLM’s new screening
LiteLLM adds configurable chatbot screening. Decide when to log or block, plan for screening failures, and account for costs and coverage limits.