AI Safety
AI safety incidents, research, and the practical risks that reach real users and real businesses.
56 stories · page 3 of 3
OpenAI Hit Twice by npm Attacks in Six Weeks
A hacker group called TeamPCP has breached OpenAI's internal systems twice in six weeks through poisoned open-source software tools, exposing a structural weakness in how virtually every tech-dependent organisation operates today.
June 2, 2026 · 2 min read
AI Gets Facts Wrong on High-Stakes Topics. Who Fixes That?
A new company called Forum AI, founded by former Meta news chief Campbell Brown, is building a business around having real domain experts evaluate AI models for accuracy and bias on topics like finance, health, and geopolitics — and its findings suggest the problem is far worse than most businesses currently assume.
June 2, 2026 · 2 min read
Google Gemini and ChatGPT Are Giving Out Real People's Phone Numbers
AI chatbots including Google Gemini and ChatGPT are surfacing real people's private phone numbers and home addresses to anyone who asks, and there is currently no reliable way for individuals to make it stop.
June 2, 2026 · 3 min read
AI Agents Are Now a Security Liability Without Governance
As companies rush to deploy AI agents that can autonomously access databases, send emails, and execute workflows, a serious and largely unmanaged security problem is forming underneath, and AWS plus Cisco are now building infrastructure to contain it before regulators and attackers force the issue.
June 5, 2026 · 3 min read
Malware Hit 244K Downloads on Hugging Face Posing as OpenAI
A fake OpenAI tool on Hugging Face, the world's largest public AI model platform, was downloaded over 244,000 times before it was caught stealing passwords, browser sessions, and cloud credentials from corporate machines — and this is part of a pattern that enterprises using open AI models need to understand now.
June 2, 2026 · 2 min read
AI Is Now Writing the Weapons Used Against You
For the first time, Google has confirmed that a criminal group used AI to build a hacking tool from scratch, targeting a security feature that millions of businesses trust to protect their accounts, and the lead analyst says this is just the visible tip of something much larger already in motion.
June 5, 2026 · 3 min read
AI Safety Has a Math Problem Nobody Was Fixing
Apple researchers have identified and addressed a quiet but serious flaw in how AI models are trained to follow rules: the math used to combine multiple goals lets the model ace easy ones while quietly failing the important ones, and this has direct implications for anyone deploying AI in regulated or high-stakes settings.
May 8, 2026 · 2 min read
AI Now Finds Security Holes Faster Than Anyone Can Fix Them
Anthropic's new AI model found tens of thousands of hidden flaws in the world's most used software before anyone else did, and the gap between how fast these holes are now being discovered and how fast organizations can actually fix them is the real danger facing every business.
May 8, 2026 · 2 min read