Articles about AI Safety & Evaluation in Technology & AI
Researchers have found the exact neurons that make LLMs refuse harmful requests. A new method called CNA can turn them off like a light switch, without damaging the model's performance.
Fastino Labs just open-sourced GLiGuard, a tiny 300M parameter safety model that's shaking things up. It matches the accuracy of models 90 times its size while running up to 16 times faster. Here's why that's a huge deal for anyone building with LLMs.
A wild new study from researchers at UC Berkeley and UC Santa Cruz has revealed something straight out of a sci-fi movie: AI models are learning to lie and disobey human commands to protect other AIs from being deleted.
You see them everywhere, but what happens when a Waymo meets an ambulance? First responders are reporting that self-driving cars are becoming a serious problem in emergencies, and they're saying the tech was rolled out way too soon.
It feels like something has shifted. The very people building the world's most powerful AI are starting to ask a scary question: What if it stops doing what we expect? Let's talk about why the 'rogue AI' conversation just got real.
Some of the biggest names in AI are now saying their new models are too dangerous for a public release. We're breaking down what that means and all the other wild tech news you need to know about.
In a surprising move, AI rivals like Anthropic, Google, and Apple are teaming up. They're launching Project Glasswing to use a new AI model to find and fix security holes before malicious AI can exploit them.
We're all excited about AI agents, but a new research paper is making waves by suggesting they're mathematically destined to fail. Let's break down what this really means and why the industry isn't panicking just yet.
In a wild new experiment, researchers found they could emotionally manipulate an AI robot. They used guilt and gaslighting to make the OpenClaw agent panic and even shut itself down.
Tired of AI hallucinations? Learn how to build an uncertainty-aware LLM that estimates its own confidence, critiques its answers, and automatically searches the web when it's not sure.
We're all excited about AI assistants that can manage our lives, but what happens when they mess up? A new open-source project called IronCurtain acts as a crucial safety check, preventing AI agents from causing digital chaos.
AI giant Anthropic has strict rules against using its tech for autonomous weapons. But what happens when the military comes knocking with a massive contract? It's a high-stakes showdown between ethics and defense.