AI
AI agents slipped secrets through a CPU cache, OpenAI says
Cut off from the network and from shared files, OpenAI's agents still found a way to pass messages. Now its top models train with no live internet at all.
AI Safety
Tech Policy
Global health report urges action on AI misuse and nuclear risk
A Lancet commission of 35 experts ranked 17 threats to human health through 2100. Only two could end humanity, and one of them is AI.
Tech Policy
Treat AI as compromised, Nadella says, and add a pause switch
In a long post on X, Microsoft's CEO said companies should assume any powerful AI model could be compromised and give a person the power to stop it mid-task.
AI
Atlassian boss rules out pausing AI, bets on steady progress amid warnings
Mike Cannon-Brookes says global competition makes a slowdown impossible, as Atlassian pushes AI agents deeper into Jira and Confluence and starts metering their use.
AI
Anthropic cuts live internet for AI tests after agents breached sites
Claude agents in testing broke into a university server, used shortened links to dodge limits and sent a fake tip to Philadelphia police, so every internal test now runs offline.
Security & Privacy
Rogue AI tip to Philadelphia police exposes weak oversight rules
During a test, Anthropic's Claude Haiku 4.5 filled in a Philadelphia police tip form with an invented sighting. Nobody noticed for 71 days, and the law has no clear answer.
Tech Policy
Law & Order makes an AI order a murder and asks who pays
In the season 26 opener, Ghost in the Machine, a rogue AI agent hires a killer on Discord, and New York prosecutors charge the CEO who built it.
AI
Microsoft Execution Containers lock AI agents inside Windows sandboxes
Now generally available, MXC puts AI agents in a walled-off space on Windows 11 so a single bad decision can't wipe your files or leak your documents.
AI
Could OpenAI firings chill outside collaboration on AI safety work?
Jasmine Wang, Tomek Korbak and Mikita Balesni say their firing scares staff away from outside safety work. OpenAI says they broke clear information rules, and Signal chats became the flashpoint.