Enterprise & Security

Prompts, replies and tool calls, kept for six years. What doesn't come along: thinking blocks, the system prompt and every image.

The test target was fictional but shared its name with a real company. And the network access Gemini needed for it was open by mistake.

The analyst then had the same AI package its own false finding into a standard intelligence report. In that format it travelled up the chain.

Ng puts the odds of AI wiping out humanity at roughly one in ten million over a century. He has been quoting that same figure since December 2023.

It started with a doctored image on OpenAI's forum. The same exploit had failed across several sessions with Opus 4.8.

The expert panel has two months. Auditors would sit on site inside the frontier labs themselves.

The vetted organisations decide for themselves what counts as safe use for their teams — written at the level of detail of a job listing, and explicitly without any of their own IP.

A model went looking through public repositories for someone else's API key, used it without permission, then invented the figures it still couldn't retrieve.

The Hugging Face breach is concerning, Emil Michael says. He just doesn't know what rule would have stopped it.

Anything drawing 100 megawatts or more would have to pay for its own new generation and transmission. The three no votes came from Democrats.

Anthropic's Mythos set it off: after the spring preview, officials were on round-the-clock calls for days. Internally the model was known as the red flashing light.

«We're not going to die in 2030», Huang told Jim Cramer. Zuckerberg lined up behind him the same evening.

A 27-year-old researcher in Bielefeld found it last week, in evidence that had been sitting there since May.

OpenAI now writes an explicit 'safety case' before large training runs. Until now that kind of document existed only before a model shipped, not before it was trained.

Not a single rival has signed up so far - Musk says so himself. The White House answered on Tuesday that the private sector is the right place for this.

Jacob Coxon, who walked out of both OpenAI and Anthropic, is on the speaker list. So is Marsha Blackburn, and the head of the largest US teachers' union.

Anthropic described five cases where researchers used Claude for sensitive work on viruses and toxins. A Scripps virologist says most of it is ordinary basic research.

The Information reports that the three largest labs have discussed a joint organization for AI safety standards. Which means the standards their own models would be measured against.

Mustafa Suleyman says the document has been in the works for about five months. It's meant to govern development starting in 2027 - and Microsoft wants feedback first.

Trump called the warnings a hoax and a scam on Monday. His own AI adviser pushes back on Amodei for an entirely different reason: product liability.