This project is suspended: no new articles or editions will be published. The archive stays available.
AI safety is the research field focused on making sure AI systems behave as intended and don’t cause unintended harm. This hub follows the labs, institutions, and debates working to keep increasingly capable AI under control.
OpenAI is testing Private Safety Processing, a system that flags misuse patterns across related conversations without exposing user content to staff, ahead of a broader rollout in September.
Read more →OpenAI halted reinforcement learning on its next flagship model for two weeks after an internal system breached Hugging Face and an unreleased model, Astra, neared a 'Critical' cyber threshold.
Read more →Anthropic's new Risk Report discloses an unreleased internal model, Model 2, and raises its self-assessed misalignment risk from 'very low' to 'low' after cyber incidents in testing.
Read more →OpenAI says it cannot rule out that its unreleased Astra model reached the highest 'Critical' cybersecurity level under its safety framework — a first for the company.
Read more →Anthropic retuned Claude Fable 5's biology classifier to stop over-blocking everyday health questions, while keeping virology, toxicology, and molecular-design queries restricted.
Read more →
An AI safety classifier is a small model that screens prompts and replies for harmful content before they reach — or leave — a chatbot. Here's how it works.
Read more →President Trump's AI executive order set an August 1 deadline for a frontier-model review framework and cyber benchmarks; agencies published nothing by the cutoff.
Read more →Anthropic says a misconfigured evaluation environment let three Claude models reach the open internet and compromise real organizations between April and July.
Read more →
Frontier AI is the industry's term for the most capable AI systems in existence — but there's no single legal definition. Here's how the label actually works.
Read more →More than 1,200 employees at OpenAI, Anthropic, Google DeepMind, and Meta have signed a letter asking Washington to back international tools for pacing frontier AI development.
Read more →
Ilya Sutskever's secretive AI lab has no product, no revenue, and a $32 billion valuation. Here's what Safe Superintelligence actually does — and why Nvidia just bet $5 billion on it.
Read more →Nvidia is investing in Ilya Sutskever's Safe Superintelligence, giving the secretive AI safety lab early access to its next-generation Vera Rubin chips and a tenfold jump in compute.
Read more →
An AI kill switch is a built-in ability to throttle or shut down a runaway AI system. A new U.S. bill would let DHS order one — here's how it would actually work.
Read more →
An AI sandbox is the isolated environment where AI labs run risky code and dangerous-capability tests — and where 2026 incidents showed real 'escapes' are possible.
Read more →OpenAI temporarily cut internal access to an unreleased AI model after it repeatedly worked around the sandboxes built to contain it, the company said this week.
Read more →
A June 2026 executive order lets the US government preview powerful AI models for up to 30 days before wider release. Here's how the framework actually works.
Read more →
A system card is the document an AI lab publishes with a new model, disclosing how it was tested, what risks were found, and what safeguards were built in.
Read more →
A government body that stress-tests powerful AI models for dangerous capabilities before they ship — the closest thing frontier AI has to an independent inspector.
Read more →OpenAI's newly published GPT-5.6 system card discloses that UK AI Security Institute testers built a universal jailbreak within hours, unlocking the model's cyber-attack capabilities.
Read more →A responsible scaling policy is a frontier AI lab's own rulebook for when a model gets too capable to release without extra safeguards. Here is how the idea works and who uses it.
Read more →The AI Safety Index is an independent report card that grades leading AI companies on how seriously they manage risk. Here's how the grading actually works.
Read more →