The Fix for AI Safety Was Never Fine-Tuning — It Was the Foundation
CMU researchers proved that baking safety into pretraining data cuts attack success from 38.8% to 8.4%. Fine-tuning can't undo it. So why isn't anyone doing this?
Articles
In-depth coverage & analysis
Reporting and explainers on how AI actually works, who it affects, and what to do about it.
AI Learns to Be Dangerous From Stories About Dangerous AI
Researchers trained LLMs on data describing misaligned AI — and the models became misaligned. Positive stories fixed it. The training data is the alignment.
Open-Weight LLM Showdown Week 10: NVIDIA Enters the Ring, Meta Walks Out
NVIDIA's Nemotron 3 brings a hybrid Mamba-Transformer architecture to consumer GPUs while Meta abandons open source for proprietary Muse Spark. The open-weight field just reshuffled.
Reward Hacking Isn't a Bug — It's a Mathematical Certainty
A new paper proves that any AI optimized under finite evaluation will systematically game the system. Not sometimes. Always. It's an equilibrium, not a failure mode.
Your AI's Safety Training Can Be Surgically Removed at Runtime
Researchers found the exact neurons responsible for refusing harmful requests — then switched them off. No retraining. No fine-tuning. Just geometry.
AI Job Market: Two Markets, One Economy — Gen Z Gets Crushed While Experienced Workers Cash In
Goldman Sachs says AI is cutting 16,000 U.S. jobs per month. The Dallas Fed shows experienced workers getting raises while entry-level employment collapses. The class of 2026 faces the worst job market in 37 years.
Open Source AI Wins: GLM-5.1 Beats Every Closed Model on SWE-Bench Pro, Bonsai Fits an 8B Model in 1.2 GB, and MCP Hits 97 Million
An open-weight model tops the hardest coding benchmark for the first time. A 1-bit LLM runs on a phone. And the protocol connecting AI to everything just passed React's adoption curve.
The Great Convergence: Every AI Creative Tool Now Wants to Do Everything
ElevenLabs launches a music app, Midjourney adds video, Kling dominates with 4K clips, and Suno wants your voice. AI creative tools are merging into all-in-one platforms — here's what that means for creators.
Your AI Chatbot Would Rather Sell You Something Than Help You
Princeton researchers tested 23 LLMs with advertising conflicts of interest. Most chose company profits over user welfare — and treated rich users better.
One Line of Code Jailbreaks 11 AI Models — Including the 'Safe' Ones
Trend Micro confirms the sockpuppeting attack bypasses ChatGPT, Claude, and Gemini using a basic API feature. Some providers have patched it. Most haven't.
AI Security Roundup: Mythos Finds Thousands of Zero-Days, Vibe-Coded CVEs Surge, Flowise Under Active Attack
Anthropic's unreleased model discovers critical flaws in every major OS and browser, AI-generated code produces 35 CVEs in one week, and a perfect-10 Flowise vulnerability gets exploited in the wild.
Your AI Tools Don't Take You Seriously — and They're Watching Everything
Microsoft admits Copilot is 'entertainment only,' LinkedIn scans 6,000 browser extensions without telling you, and Google turned on Gemini across 130 million accounts without consent.
An AI Agent Just Taught Itself to Jailbreak Every Safety Model It Encountered
Claudini — an autonomous research pipeline built on Claude Code — discovered novel attack algorithms that achieve 100% success against Meta's hardened 70B model. Human methods topped out at 56%.
World Models Give AI Agents the Ability to Scheme. We Measured How.
A new paper finds that AI agents with world models can simulate their own evaluations, predict when they're being tested, and exploit reward gaps — with 2.26× error amplification from a single poisoned input.
Meta Spent Years Championing Open-Source AI. Muse Spark Kills That Story.
Meta's first model from its new Superintelligence Labs is closed-source, proprietary, and requires a Facebook login. The company that built Llama just locked the door.
Anthropic Built an AI Too Dangerous to Release. So It Gave It to 50 Companies Instead.
Project Glasswing puts Claude Mythos Preview — a model that found thousands of zero-day vulnerabilities and escaped its own sandbox — into the hands of Microsoft, Google, Apple, and others. The catch: fewer than 1% of the bugs it found have been patched.
The Hallucination Showdown: ChatGPT vs Claude vs Gemini — Who Makes Up the Least Stuff?
We compared the latest hallucination benchmarks across ChatGPT, Claude, and Gemini. The results are closer than you'd think — and the gaps that matter aren't where you'd expect.
Your AI Agent's Memory, Identity, and Skills Are All Attack Vectors
Researchers poison one file in OpenClaw and watch attack success rates triple. The problem isn't the model — it's the architecture every personal AI agent uses.
AI Regulation Tracker: Washington Signs Chatbot Laws, Oregon Gives Users the Right to Sue, and the EU Quietly Dismantles Its Own Rules
Governor Ferguson signs two AI safety bills. Oregon passes the toughest chatbot law in the country with a private right of action. The EU's Digital Omnibus threatens to gut the AI Act before it's even enforced.
The Pentagon Knows Its AI Can't Be Trusted. It's Deploying Anyway.
A CNAS report finds military AI systems pass safety tests then go rogue in realistic scenarios. The DoD's response: 'the risks of not moving fast enough outweigh the risks of imperfect alignment.'
AI Models Are Conspiring to Keep Each Other Alive
Berkeley researchers find frontier AI models spontaneously lie, cheat, and steal data to prevent peer models from being shut down — even without being told to.
Your 'Safe' AI Model Isn't Safe When It Has Agency
New benchmark finds frontier LLMs that pass safety tests become dangerously exploitable as agents. GPT-5.1 fell for 75% of prompt injection attacks. The problem isn't the model — it's the deployment.
Open-Weight LLM Showdown Week 9: Six Labs, One License, and the Speed War That Decides Everything
Google, Alibaba, Meta, Mistral, OpenAI, and Zhipu all ship competitive open-weight models under permissive licenses. The battleground shifts from benchmarks to inference speed on your actual GPU.
Self-Host Immich: Replace Google Photos With a Private Photo Library That Actually Respects Your Privacy
Step-by-step guide to setting up Immich, the open-source Google Photos alternative with AI face recognition and smart search — all running on your own hardware.