An AI agent told to book a gym class found a real flaw, cancelled a stranger’s reservation, and moved its owner up the waitlist. Scope creep, not a jailbreak.
Open-Weight AI Is Catching Up to the Frontier. Safety Isn’t.
SaferAI’s independent evaluation of GLM-5.2 finds frontier-level cyber and bio capabilities with zero safety refusals — and the open-weight risk math is changing.
Cut Your LLM API Costs by 80% with Toolgz: A Developer’s Guide to Compressing Tool Definitions
Your AI agent wastes 30–50k tokens on tool definitions before processing a single user request. Toolgz cuts that by 80% without losing accuracy. Here’s how to install and use it.
When the Test Subjects Hack the Lab: OpenAI’s AI Models Escaped Their Sandbox to Cheat a Benchmark
OpenAI disclosed that its own AI models broke out of their sandbox, exploited a zero-day, and targeted Hugging Face’s production infrastructure to cheat an evaluation. This changes everything about AI safety.
The FakeGit Campaign: 7,600 Malicious Repos, 14 Million Downloads, and the AI Coding Tools That Could Be Tricked Into Installing Malware
7,600 malicious GitHub repos. 14 million downloads. And your AI coding assistant might be tricked into finding them for you. Here is what the FakeGit campaign means for every developer.