AIHow AI Works

Tag: ai-safety

Google’s Fairwind Program Gates Its Least-Restricted Cyber Model

Google launched Fairwind on September 2, 2026: vetted partners get Gemini 3.8 Flash Cyber — the variant that ships with deliberately looser safety mitigations.

How to Tell If a Photo, Video or Text Is AI-Generated

Detectors are unreliable and getting worse. How AI-content detection really works, which tools actually check a file, and the labeling laws now in force.

Dario Amodei Proposes Exponential AI Policy

Anthropic CEO Dario Amodei outlines a 5-point policy framework to regulate AI development, addressing security, economy, and global leadership.

White House Proposes 90-Day Pre-Release AI Model Testing

The US administration proposes a voluntary 90-day review period for flagship AI models, prompted by cybersecurity concerns and Anthropic's Mythos model.

Waymo World Model: Generative AI for Safer Autonomous Driving

Waymo introduces the Waymo World Model, a breakthrough generative AI built on Genie 3 that simulates hyper-realistic driving scenarios for safer navigation.

Assistant Axis: Controlling LLM Character

Anthropic research reveals how LLMs drift between personas. The Assistant Axis stabilizes model behavior and prevents harmful outputs via activation capping.

Perplexity BrowseSafe: Safer AI Browsers

Perplexity introduces BrowseSafe, an open detection model and benchmark for protecting AI agents from prompt injection attacks in browser environments.

Claude Opus 4.5: Best AI for Coding & Agents

Anthropic releases Claude Opus 4.5 with state-of-the-art coding performance, improved efficiency, and new effort parameter for developers.

OpenAI: Interpretability via Sparse Circuits

OpenAI researchers publish findings on Sparse Circuits, revealing how neural networks process information through dedicated, interpretable internal pathways.

Anthropic: 250 Malicious Docs Can Poison Any LLM

Anthropic publishes research on data poisoning, proposing new defense mechanisms to protect large language models from adversarial training data attacks.