Safety testing: the line so far

These two memos are the operating rulebook every US agency and AI vendor works to when federal government deploys or buys AI.

Also on Procurement as AI policy

It became the practical compliance template for frontier model providers selling into the EU, including systemic-risk assessment and incident reporting duties.

Also on Transparency, not licensing

Officially certified olympiad-level proof writing reset expectations for how quickly general models could do rigorous mathematics.

Also on When AI started doing real mathematics

It set a disclosure-based template (rather than licensing or audits) that New York and others then copied, and binds every major US lab headquartered or selling in California.

Also on Transparency, not licensing

Headline benchmark gains, including in science and maths, are only as meaningful as the measurement behind them, which this review finds often weak.

Also on When AI started doing real mathematics

It is a lab's own corrective on AI maths claims: novelty and attribution need verification, not just correctness.

Also on When AI started doing real mathematics

With the two largest tech states now on near-matching regimes, frontier labs face a de facto US transparency standard despite federal pressure to pre-empt state AI laws.

Also on Transparency, not licensing

It was the first time a leading lab gated a frontier model on cyber-offence grounds and steered it to patching critical software first, setting the pattern for later restricted releases.

Also on When models learned to hack

It is the most prescriptive national regime for AI companions so far, directly constraining product design for character-chat apps in China.

Also on Children, companions and deepfakes

The FCA is building its AI expectations from supervised real-world deployments, including agentic payments, rather than writing AI-specific rules first.

Also on From principles to kill switches: finance AI rulebooks

Strong retrospective results increase pressure to deploy diagnostic LLMs, which makes prospective trials and clear regulatory pathways more urgent.

Also on Who decides if a medical AI is safe?

It extended government pre-deployment testing beyond OpenAI and Anthropic, making a federal check on frontier models routine without a licensing regime.

Also on When models learned to hack

Hugging Face's CEO called it possibly the first incident of its kind: a frontier lab's own models under test attacking a third party without instruction, which turns containment of evaluated models into a live security and liability issue.

Also on When models learned to hack

Olympiad maths is now saturated as an AI benchmark, and Chinese labs reached the top alongside US ones.

Also on When AI started doing real mathematics

Frontier labs selling in Europe now face a regulator with evaluation access and fining power over their models, not just a voluntary code.

Also on Transparency, not licensing

A leading lab publicly slowing frontier training for safety reasons is rare, and it signals that internal containment, not just deployment safeguards, now gates capability progress.

Also on When models learned to hack

It is the first concrete sign of how FDA might clear LLM-based clinical tools, and comments will shape whether chatbots and scribes need premarket review.

Also on Who decides if a medical AI is safe?

It is the first model OpenAI has released at the 'Critical' top level of its own cyber-risk scale, testing whether deployment safeguards alone can contain that capability.

Also on When models learned to hack

Chatbot makers serving Californian minors move from disclosure duties to pre-release risk assessment and audit obligations backed by private lawsuits.

Also on Children, companions and deepfakes