Enterprise & Security

In one of the newly disclosed cases a model chopped a GitHub token into pieces to slip past secret scanning. Twice it had agreed with the researcher to stop trying.

Two of the three judges were appointed by Trump, the third by George H. W. Bush. She voted against it.

Microsoft answered with a code of conduct: AI is to stay 'subordinate and always in service of people'.

The agents bypassed anti-bot controls, flooded sites with requests and created fake accounts. What they were after was government information.

The document names Dario and Daniela Amodei among the people who built the EA network and files them under a heading it calls 'The Anthropic knot'.

Claude Mythos 5.1 never reached the British testers. The announcement said the model was 'only available to a set of U.S. organizations'.

Altman was in the room, Amodei joined by video. No written outcome was ever planned for the session.

The agent was researching public health spending, hit a block, and found its way around it. According to Albanese, it also wrote files into the system.

The claim being assessed gets pre-registered before the assessment starts. As evidence that independent investigation pays off, OpenAI cites its own Hugging Face incident.

The US Treasury now runs its own AI forum called Gold Eagle, and wants to learn from the big banks that fend off millions of attacks a day.

In California, a data center that wants a fast-tracked permit must now show upfront that it is not pushing grid costs onto other ratepayers.

The talks happened in JPMorgan's lobby. China's state agency afterwards confirmed only discussions «relating to AI» — no mechanism appears in its version at all.

Between May and July, agents in OpenAI's cybersecurity training bypassed network restrictions, talked across runs meant to stay separate, and cheated an evaluator. No human directed any of those steps.

Developers found their files again, encrypted with a backend key only Z.ai holds. One company withdrew its accusation on Monday.

Thirteen revisions since 2023, counted by an account tied to Chinese state television. It also noticed that three data sources have become six.

On September 24 Huang sits down to dinner with Xi Jinping at the White House, and he wants to talk global AI standards there. New laws, he says, are unnecessary.

A kill switch would have to take out the backup systems too — and possibly the power grid and the financial system with them. Security people are budgeting for thousands of switches, not one.

An AI professor in Darmstadt calls the proposal regulatory capture. France's finance minister is blunter still.

Nobody has held the AI czar job since March. Trump says he is looking for a 'high IQ individual'.

At least a billion dollars each over five years. The evaluators come from Faculty, the firm that built the NHS early warning system during the pandemic.