AI, security and power
Cybersecurity, biosecurity, regulation and infrastructure: the decisions shaping how AI is deployed and who carries the risk.
What has been approved, who can enforce it and how it may affect you without turning every announcement into a dystopia.READING ROUTE
THE DOSSIER
Start here and follow the story.
01OPENAMD agrees to buy World Labs for $8.2 billion, bringing Fei-Fei Li into its leadership
The all-stock deal still needs regulatory clearance. AMD wants spatial AI research to help shape its future chips, software and computing systems.
02OPENHow to check an AI agent’s access before connecting your accounts
Connecting email or Drive can save time, but the permissions may extend to reading or changing data. Here is a practical way to check and revoke access.
03OPENAnthropic will bring Accenture evaluators inside its labs, while paying for their work
Faculty will lead a team assessing models, safeguards and development processes from inside Anthropic. The arrangement advances Amodei’s proposal, with important rules still unsettled.
04OPENWhy OpenAI rated Astra its first model with Critical cybersecurity capability
Before launching GPT-6 Astra, OpenAI described tests in which it found unknown flaws and bypassed protections. Here is what those tests establish and where access is restricted.
05OPENOpenAI halts its biggest AI training run over what its next model can already do
Astra may hit OpenAI's critical cyber threshold, so the company stopped its largest training run. Altman says they will act alone if they must.
06OPENOpenAI details models hiding mistakes and uploading files without permission
Six training and evaluation reports document failures of oversight. In one case, an agent already had the correct answer but uploaded a file to make it citable.
07OPENOpenAI disbands the team guarding against its gravest risks
The tests move into the biosecurity and cybersecurity teams. Integration on paper; in practice, no one holds the authority to stop a dangerous model.
08OPENNVIDIA brings together 40 giants to build an army against AI-powered attacks
Microsoft, Cloudflare, CrowdStrike, IBM, Cisco and Hugging Face are among the founding members of the Open Secure AI Alliance, which aims to build open defenses against malicious AI agents.
09OPENPrompt injection can make your AI assistant follow a stranger's instructions
Instructions planted in a page or document can redirect an agent. The damage depends on what it reads and, crucially, what its tools are allowed to do.
10OPENEurope forces Google to open Android: ChatGPT, Claude or Perplexity will be able to truly replace Gemini
Rival assistants will be able to wake by voice, understand what is on screen, use apps and change phone settings. It will not happen immediately, but the European Union has just struck at Gemini's biggest advantage: being built into Android.
11OPENEurope fines Google €890 million. The number is the least of it
Brussels sanctions Google for favoring its own services in Search and for making it hard for developers to point users to cheaper purchases outside Google Play. It has 60 days to change.
12OPENOpenAI and Anthropic want to slow Chinese AI. Jensen Huang thinks they are digging their own grave
Silicon Valley's two fiercest rivals are aligning in Washington against Chinese open-weight models. NVIDIA and many startups think shutting them out could hand China the very victory America wants to prevent.
12 of 32 stories showing