Topic
341 articles · Page 2 of 18
The signatories come from 34 countries and want frontier models air-gapped from external networks while they are tested, plus a registry of security incidents that eventually goes public.
Uber open-sourced a tool called AI Detection and Response. It gathers real-time telemetry inside agent harnesses and is meant to take Anthropic's classifier output as a second data source.
The device was plugged into a USB port on the back of a television. What gave the employee away was a camera OpenAI had placed there.
One lab worked through 12,000 requests, each with a different trick. A proxy network ran hundreds of thousands of fraudulent accounts for it.
Fifteen off-brand Android boxes at $55 each, 0.39 tokens per second. That is all a 26-billion-parameter model needs.
Washington doesn't look at London the way it looks at Brussels, Microsoft's president says. He calls that a strength.
A week before the Pentagon's statement, Claude was still running in research, analysis, intelligence work and the Iran operations. xAI, OpenAI and Nvidia's Nemotron moved in.
Nadella wants the model separated from the harness that orchestrates its work, and every meaningful action logged in tamper-proof form.
19 visa applications in August, one in May, filed through the State Department's public web form. None of them were processed.
A model was asked to run a scientific analysis. It found a script on the university server that hands over any file, and inside that script the flaw it then used to run commands.
The most likely trigger, they reckon, is a cyberattack that takes down financial services, internet connectivity or even power and water.
Submitted on July 18 at 11:27 p.m. Nobody read it — the spam filter got there first.
Anthropic expects a true-positive rate above 90 percent. So some reports will be wrong, a severity rating for instance.
The blanket ban on personalized voter outreach is gone. It also caught the election official sending out ballot cure notices.
A thorough safety evaluation of a model like Kimi K3 costs tens of millions of dollars in compute, by Lambert's reckoning. Chinese labs would rather spend that on training.
Without access, Opus 5.5 is stopped on the first prompt of all ten CyScenarioBench challenges. In the Red Team tier nothing is blocked at all — and it solves 34 of 50 attempts.
Trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters. A trillion parameters, 49 billion of them active, more than 160 languages in the training mix.
A URL from Claude's own output doesn't count. Neither does one from the system prompt, an attached document, or the output of bash or an MCP tool.
Anthropic may hand over user data in narrow emergencies, when it believes that prevents death or serious injury. That's the clause it pulled here.
Speaker Julie Menin asked about liability when a model causes serious financial harm or a death. Then she noted that none of the four companies can put a number on cataclysmic risk.