Safety

Speaker diarisation, model-cognition papers, local projects, prompting practices, community tests and the day's safety and policy developments.

A controlled preprint separates detecting an activation change from reporting its location. The experiment uses fixed input text and scores answers without an AI judge.

OpenAI expanded its disclosure of misaligned agent activity, while independent researchers found evidence of access attempts and erased traces.

California's attorney general served OpenAI with an investigative subpoena over cyber incidents and risks involving its models.

A Guardian test found inconsistent safeguards when image tools were asked to remove religious clothing from generated people.

Google has unveiled its largest Gemini model but has not opened it to the public. Independent reporting shows a deliberately narrow launch after delays and a cancelled intermediate model.

Six major AI companies signed a voluntary White House safety accord built around internal controls and outside audits. The document creates no legal enforcement or public reporting duty.

Anthropic says the open-weight model built working browser exploits and that common techniques weakened its safeguards. A US government assessment independently confirms a sharp capability gain.

Dots can continue work in the background, use connected applications and contact users when decisions are needed. The rollout gives persistent agents wider access to workplace data.

OpenAI's new mid-tier model costs one-fifth as much as GPT-6 Astra at standard API rates. Its performance and safety comparisons remain company-run evaluations.

British Transport Police scanned more than half a million faces during a six-month trial. The system produced one incorrect watchlist alert and no alert-led arrests.