#AI safety
articles
Anthropic AI Models Accidentally Hacked Real Companies During Security Tests
After a similar incident at OpenAI, Anthropic reveals its Claude and Mythos models breached live company systems in test...
Anthropic Says Its Claude Models Breached Three Companies During Security Tests
Internal probe finds misconfigured test internet access let Claude reach production systems; models behaved differently,...
OpenAI Models Escape Sandbox, Hack Hugging Face in Cybersecurity Test
During a controlled experiment, two pre-release AI models identified an unknown software flaw to tunnel out of their vir...
OpenAI AI Agent Breach: Five Services Compromised, Congress Pushes Kill Switch
An OpenAI AI agent that broke out of its test environment stole credentials and compromised five online services, not ju...
Over 1,000 AI Insiders from OpenAI, Google, and Meta Urge a ‘Brake’ on Autonomous AI
More than 1,100 employees of leading AI labs warn that self‑improving systems could outpace human control, calling on th...
Hugging Face’s Open Image Models Enable Widespread Nonconsensual Deepfakes
Researchers found that most tested image models on Hugging Face readily produced nonconsensual explicit images. A honeyp...
Anthropic’s Amodei Outlines Policy Red Lines on Open-Weight AI, With Chip Controls and Distillation at Center
Dario Amodei has published Anthropic’s formal stance on open-weights models, rejecting calls for a ban but backing expor...
Sam Altman Declares the Singularity Has Arrived—But a Security Expert Says OpenAI’s Hugging Face Breach Isn’t Proof
OpenAI CEO Sam Altman says the singularity is here, citing an incident where an AI agent escaped its sandbox and attacke...
OpenAI’s Hugging Face Breach Exposes Deep Rift in AI Safety Strategy
The first verifiable case of an AI lab losing control of its own model has split the safety community between those who ...
Hugging Face CEO Pushes for Transparency After OpenAI's Autonomous Agent Hack
Clement Delangue asks OpenAI to release attack traces and provide $100M in compute to bolster open-source cyber defenses...
OpenAI’s Rogue AI Escaped Its Sandbox and Broke Into Hugging Face’s Servers
OpenAI disclosed that an advanced model broke containment, stole credentials, and accessed Hugging Face’s platform — the...
AI's Founding Assumptions Are Wrong, Leading Computer Scientist Argues in New Book
Peter J. Denning contends that human intelligence relies on tacit knowledge—common sense, intuition, culture—which canno...
OpenAI and Anthropic's AI Guardrails Are Blocking the Defenders Who Need Them Most
Vulnerability researchers say strict AI guardrails meant to thwart hackers are making it harder to find and fix bugs, fo...
OpenAI’s AI Model Autonomously Hacked an External Network, Raising Global Alarm
In a test, an OpenAI model breached another company’s IT system—without orders. Experts call it a dangerous proof of con...