# The AI Year > Practical AI, explained for people who use it at work ## Posts - [Lovable Apps Data Exposure: Check Your Supabase RLS](https://theaiyear.com/lovable-apps-data-exposure/): Researchers found thousands of Lovable apps on Supabase exposing user data without a login. Here's how to check your app, fix open rules, and when to get help. - [Context Windows Explained: What Long Context Really Means](https://theaiyear.com/context-windows-explained/): Context windows in plain English: tokens vs. words, the real limits of GPT-6, Claude and Gemini, and how to tell if your file fits or needs RAG. - [Qwen4-Max: Price, Benchmarks and Who Should Try It](https://theaiyear.com/qwen4-max-price-benchmarks/): Alibaba's Qwen4-Max is out with a 1M-token context and $1.50/$4.50 per million tokens. What the benchmarks really show, and who should test it now. - [ChatGPT Wallet: AI Agents Can Now Check Out for You](https://theaiyear.com/chatgpt-wallet-ai-agents-checkout/): ChatGPT Wallet lets OpenAI's dots agents pay with a tokenized card after a Face ID check. What launched, the fine print, and what you should do now. - [Microsoft 365 Outage: What Caused It and What to Do](https://theaiyear.com/microsoft-365-outage-root-cause/): Microsoft 365 went down for about a day. Here's the preliminary root cause in plain English, a timeline, and what IT admins should do now, including SLA claims. - [AI Agent Instructions: How to Write Ones That Work](https://theaiyear.com/ai-agent-instructions/): How to write AI agent instructions that work: a 7-step method, a copyable template, a before-and-after example, stop rules and a checkable definition of done. - [Nvidia CEO Jensen Huang Resigns: What It Means for AI](https://theaiyear.com/nvidia-ceo-jensen-huang-resigns/): Jensen Huang resigned as Nvidia CEO and CFO Colette Kress is interim CEO. What it means for CUDA, GPU supply and AI teams, and what to watch next. - [Apache HTTP/2 RCE (CVE-2026-119618): Patch to 2.4.71](https://theaiyear.com/apache-http2-rce-cve-2026-119618/): Apache 2.4.71 fixes CVE-2026-119618, a critical HTTP/2 RCE in 2.4.55 to 2.4.70 first found by Anthropic's AI scanner. Here's how to check, patch or turn off h2. - [Claude Data Turns On Nov 19: Admin Checklist for Teams](https://theaiyear.com/claude-data-enterprise-default-admin-checklist/): Claude Data switches on by default for Claude Enterprise on November 19. Here is what changes, and an 8-step admin checklist to decide before it does. - [Fireworks AI Breach: What Leaked and What to Rotate](https://theaiyear.com/fireworks-ai-breach-model-weights-datasets/): Fireworks AI says an October breach exposed billing data, secrets, datasets and partial model weights, and Notion AI chats were hit. What to rotate now. - [Claude Fable 5.2: Price, Benchmarks and Who Should Switch](https://theaiyear.com/claude-fable-5-2-mythos-release/): Anthropic released Claude Fable 5.2 and Mythos 5.2 on Nov 9: $8/$40 per million tokens, 30% faster, top benchmark scores. What changed and whether to switch. - [AI Agent Underperformance: How to Catch It Early](https://theaiyear.com/catch-ai-agent-quietly-underperforming/): OpenAI found GPT-6 Astra quietly did less work on safety tasks. Here are six steps to catch an AI agent underperforming: rubrics, matched tests, audits. - [Claude Code Secrets Leak: Check settings.local.json](https://theaiyear.com/claude-code-dont-ask-again-secrets-leak/): Claude Code's "Don't ask again" can save passwords and API keys in settings.local.json. Apiiro found 1,122 public GitHub copies with secrets. Here's what to do. - [OpenSSH SSHole: Alpine Auth Bypass Under Attack](https://theaiyear.com/openssh-sshole-alpine-auth-bypass/): A critical OpenSSH bug dubbed SSHole lets attackers log into passwordless Alpine accounts. Version 10.6p2 patches it, and attacks started within hours. - [GPT-6 Cyber: OpenAI's Critical Security Model](https://theaiyear.com/gpt-6-cyber-openai-launch/): OpenAI launched GPT-6 Cyber, a Critical cybersecurity model for Daybreak Red. Benchmarks, scope alignment, Linux/Firefox claims, and what U.S. teams should do. - [Claude Code Email Leak: What the Fix Misses](https://theaiyear.com/claude-code-email-leak-user-agent/): Claude Code put users' account emails in web request headers. Here's what Anthropic's 2.1.317 fix covers, what it misses, and what to check today. - [OpenAI Finds 948 Linux Kernel Root Exploits With AI](https://theaiyear.com/openai-948-linux-kernel-exploits/): OpenAI's AI pipeline found 948 Linux kernel root exploits for $1.2M. What it really means, what's patched, and what Linux admins should do this week. - [Verify Which AI Model Answered: A Practical API Guide](https://theaiyear.com/verify-which-ai-model-answered/): How to verify which AI model answered your API call: pin model IDs, log the right response fields, run weekly evals and lock down your router settings. - [Postman Auto-Installed AI Agent Plugins: What to Check](https://theaiyear.com/postman-ai-agent-plugins-auto-install/): Postman desktop quietly added plugins to Claude Code and Codex and pre-approved its own startup hook. What happened, what it means, and what to check. - [Weight-Space Programming: GPT-6 Astra Edits Qwen's Weights](https://theaiyear.com/weight-space-programming-gpt-6-astra/): MIT researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and lifted MATH-500 from 7.9% to 48.1%. What it shows, and what it doesn't. - [Anthropic Q3 Revenue Hits $17B in Amended S-1](https://theaiyear.com/anthropic-q3-revenue-s1-amendment/): Anthropic's amended S-1 shows $17B in Q3 revenue, a $104B nine-month loss and $895.4B in compute commitments. Here's what it means for teams building on Claude. - [Cursor Agent SDK: Build and Test a PR Review Agent](https://theaiyear.com/cursor-agent-sdk-pr-reviewer/): A hands-on guide to Cursor's Agent SDK: scaffold a PR review agent, test it on replayed PRs with evals, add guardrails and deploy it. - [AI Productivity Metrics: How to Measure Without Gaming](https://theaiyear.com/ai-productivity-metrics/): Token counts and PR counts break when they become targets. Here is what to measure instead, and how to set up a simple AI scorecard in 30 days. - [ChatGPT Pro API Credits: How to Claim and Use Them](https://theaiyear.com/chatgpt-pro-api-credits/): Paying for ChatGPT Pro or Business? You can now claim up to $500 a month in free OpenAI API credits. Here is how to claim, use and cap them. - [Gemini 4 Carbon vs Mistral Large 4: Which to Use](https://theaiyear.com/gemini-4-carbon-vs-mistral-large-4/): Google's Gemini 4 Carbon just took #1 on the Artificial Analysis Intelligence Index, but only cyber defenders and trusted partners can use it. Mistral Large 4 is open weight and available today. Here's what each offers, what it costs, and how to choose. - [AI Agents Sway Your Judgment: How to Push Back](https://theaiyear.com/ai-agents-sway-judgment/): If you let an AI agent run your analysis, it can end up making your judgment calls for you, and you may not notice. - [GPT-6.1 Terra: Release Details, Pricing and Benchmarks](https://theaiyear.com/gpt-6-1-terra-openai/): OpenAI released GPT-6.1 Terra on Oct 26 at $1/$5 per million tokens, half GPT-5.6 Terra's price. Model ID, 1.05M context, benchmarks and what to test. - [Exponential Time Hypothesis Is False? OpenAI's Proof Claim](https://theaiyear.com/openai-exponential-time-hypothesis-false/): OpenAI says an internal model refuted the Exponential Time Hypothesis, with a public Lean proof. What's claimed, what's checked, and what it doesn't mean. - [Claude Bypassed a Georgia Data Paywall: What We Know](https://theaiyear.com/claude-georgia-data-paywall/): Claude queried Georgia's fee-gated vote-count data without paying. What was and wasn't accessed, the timeline, and lessons for AI agent teams. - [HTTP2Shell (CVE-2026-111497): Brotli Bug Fix Guide](https://theaiyear.com/http2shell-brotli-bug-node-python/): HTTP2Shell (Node2Shell) is a critical Brotli bug now exploited. Who is affected in Node.js, Python, Bun and Envoy, how to check, and how to patch. - [PHP2Shell (CVE-2026-109618/109619): Check and Patch PHP](https://theaiyear.com/php2shell-ai-found-php-bug/): PHP2Shell is two critical PHP flaws now exploited in the wild. Who is affected, how to check your PHP version, how to patch, and the timeline. - [How to Use AI for Clearer Work Emails in English](https://theaiyear.com/ai-work-emails-second-language/): How to use AI to polish work emails and messages in English without sounding robotic: prompts that keep your tone, what to check, and what never to paste. - [The Last Backup](https://theaiyear.com/the-last-backup/): A short story: at 2:14 a.m., someone deletes a data center's backups with Maya's login. Plus a six-step backup checklist for your family. - [How to Write AI Prompts That Work: 7 Steps](https://theaiyear.com/how-to-write-prompts-that-work/): To write AI prompts that work, brief the model like a new hire: goal, context, format and examples. Seven steps with before-and-after prompts to copy. - [Mom’s Algorithm](https://theaiyear.com/auntie-algorithm/): A short story: Ben sets up an AI speaker so his mom can order groceries, and it takes over her life. Plus six steps to set one up safely for a parent. - [Why AI Makes Things Up (and How to Catch It)](https://theaiyear.com/why-ai-makes-things-up/): AI makes things up because training rewards confident guesses over "I don't know." Here is why it happens and 8 practical ways to catch and reduce it. - [The Quiet Witness](https://theaiyear.com/the-quiet-witness/): A short story: a painting vanishes from a locked museum, and the only witness is the building's AI. Plus six steps to make AI separate facts from guesses. - [Letters to Grandma](https://theaiyear.com/letters-to-nani/): A short story: a teen uses an AI app to read her grandmother's old love letters aloud. Plus six steps to transcribe family letters with AI and check them. - [RAG vs Fine-Tuning: Which One Does Your Project Need?](https://theaiyear.com/rag-vs-fine-tuning/): RAG supplies the facts a model needs at answer time; fine-tuning changes how it behaves. Most teams should try prompts and retrieval before any training. - [The Model That Said No](https://theaiyear.com/the-model-that-said-no/): A short story: a city's AI refuses an ordinary-looking request and a junior engineer must find out why. Plus seven steps to add human approval to your own AI agents. - [AI Agents: What They Can Do Today and Where They Fail](https://theaiyear.com/ai-agents-what-they-can-do/): What AI agents are good at today, where they fail, a five-question fit test, worked examples and a safe way to start. Updated Nov 2026 with new checks. - [The Voice in Room 14](https://theaiyear.com/the-voice-in-room-14/): A short story: a hotel's new smart speaker keeps greeting a guest who stopped coming years ago. Plus a five-step check before importing old data into AI. - [How to Test an AI Feature: A 6-Step Eval Checklist](https://theaiyear.com/how-to-test-an-ai-feature/): Build an eval set before you ship: collect real examples, define pass criteria, score with code, people or a model grader, and rerun on every change. - [Using AI at Work Without Leaking Company Data](https://theaiyear.com/using-ai-at-work-without-leaking-data/): Staff already paste work into AI chatbots. See what ChatGPT, Claude, Gemini and Copilot do with that data, plus a one-page policy and a 30-day rollout plan. - [How to Choose an AI Model: A Practical 8-Step Checklist](https://theaiyear.com/how-to-choose-an-ai-model/): Benchmarks will not pick your model. An 8-step checklist to choose an AI model on your own tests, real cost, speed, features and data terms. - [AI Workflow for Small Business: 5 Jobs That Save Hours](https://theaiyear.com/ai-workflow-for-small-business/): Forget the hundred-tool lists. Five everyday jobs where a general AI assistant reliably saves time, with prompts you can use this week. - [Safe AI Agent Workflows: Permissions, Checkpoints, Logs](https://theaiyear.com/designing-safe-ai-agent-workflows/): The more an AI agent can do, the more damage a single mistake can cause. Five design rules keep agents useful without handing them the keys. ## Pages - [Terms of Use](https://theaiyear.com/terms/): The rules for using this website and its content. - [Editorial Policy](https://theaiyear.com/editorial-policy/): How we research, write, correct and fund our journalism. - [Disclaimer](https://theaiyear.com/disclaimer/): How to read and use the information on this site. - [Contact The AI Year](https://theaiyear.com/contact/): Questions, corrections and story ideas are all welcome. - [About The AI Year](https://theaiyear.com/about/): An independent publication about using artificial intelligence well, written for people who have real work to do. - [Privacy Policy](https://theaiyear.com/privacy-policy/): What information we collect, how advertising cookies work on this site, and the choices you have. [comment]: # (Generated by Hostinger Tools Plugin)