AI news for Friday, November 13

North Korea's FaceStealer worked through video calls from a friend's real account. The face on screen could be a recording or a deepfake. On November 12 the Justice Department said it had seized the cloud servers behind it, along with more than 900 webcam videos of the people it targeted, and the FBI puts the take at more than $400 million in crypto over 18 months.

Four nonprofits want a judge to stop Anthropic, Google, Meta, OpenAI and SpaceXSI from pushing their most advanced models any further. They sued in San Francisco the same day. Their argument is the executives' own warnings (years of them, about catastrophe) run through a California consumer law: you said it could happen, so a court should take you at your word.

Mirendil raised $1 billion. It only came out of stealth in June, and its whole pitch is a model it says one researcher pretrained with agents writing over 99% of the code.

Use AG2's SandboxShellTool? Update. A newline let agents held to safe shell commands run anything at all. Claude Code 2.1.325 fixes a security issue in how it decodes zstd-compressed responses. And OpenAI's new ChatGPT Wallet lets its agents pay for things once you approve each purchase.