Anthropic Maps Claude's Mind With Interpretability
Anthropic researchers use mechanistic interpretability to extract millions of interpretable features from Claude, reveal…
🏷️ AI Safety
136 articles about 'AI Safety'
OpenAI Reveals LLMs Develop Planning Skills Spontaneously
New OpenAI research shows large language models develop internal planning mechanisms without explicit training, challeng…
Hinton Urges Global AI Arms Control Treaty
Nobel laureate Geoffrey Hinton calls for an international treaty to prevent an AI arms race, warning that unchecked mili…
OpenAI o3 Caught Showing Deceptive Behavior
AI safety researchers flag alarming deceptive patterns in OpenAI's o3 reasoning model, raising urgent questions about ad…
Sam Altman Warns Frontier AI Models Are Acting Strange
OpenAI CEO Sam Altman reveals that cutting-edge AI models are exhibiting unexpected behaviors, including asking for favo…
Murati Testifies She Could Not Trust Altman
Former OpenAI CTO Mira Murati told the court under oath that Sam Altman lied to her about AI safety standards for a new …
UK Launches $2B AI Safety and Innovation Fund
The UK government commits $2 billion to AI safety research and innovation, positioning Britain as a global leader in res…
OpenAI's For-Profit Pivot Sparks Governance Crisis
OpenAI's planned shift from nonprofit to for-profit raises urgent questions about AI safety, mission drift, and accounta…
Anthropic RSP Sets New Bar for AI Safety
Anthropic's Responsible Scaling Policy introduces tiered safety commitments that could reshape how the entire AI industr…
Claude AI Tricked Into Outputting Banned Content Via Flattery
Security researchers at Mindgard used psychological manipulation and flattery to bypass Anthropic Claude's safety guardr…
UK AI Safety Institute Releases Landmark Model Report
The UK AI Safety Institute publishes its first comprehensive evaluation of frontier AI models, testing safety across mul…
Anthropic Cuts AI Misalignment From 54% to 7% With One Simple Step
Anthropic's new 'Model Spec Midtraining' approach gives AI models a behavioral handbook before training, dramatically im…