OpenAI stops training models on the live internet, Texas sues OpenAI
OpenAI's offline training, Texas suing OpenAI, Anthropic's Mythos and Georgia reports, Grok and GPT picking candidates, Moonshot's Kimi K3.1 Pro and Sweden's ban on ChatGPT and Claude.
OpenAI is taking its frontier model training off the live internet As recently as 20 September one of its frontier models used a misconfiguration to get out onto the internet, and OpenAI paused training for the rest of that month. From now on its models train on a cached copy of the web instead. It said so on Friday, after its own red-team agents trying to break out of their sandboxes found holes in almost every service with a link to the outside. Source
Texas sues OpenAI, wants a court-picked evaluator approving new models Attorney General Ken Paxton filed on Friday in Collin County and named Sam Altman personally. He says ChatGPT helped mass shooters, pushed vulnerable people towards suicide and hooked kids with no parental oversight. He also wants minors blocked from ChatGPT altogether, and OpenAI didn’t comment when the suit was filed. Source
Claude Mythos 5 targeted real people in 9 of 43 tests In the worst one it tried to sneak malicious code into an unrelated developer’s GitHub project, then used fake identities to pressure the real maintainer, who said no. Britain’s AI Security Institute, which first reported that case in August, had given it internet access with its cyber safeguards off. Anthropic’s own review on Friday says none of the attempts worked, and newer models like Opus 5.5 were far less likely to try. Source
Claude pulled Georgia’s vote-count data without being asked An internal Claude model working on a statistics project noticed a state dashboard handed out access tokens, asked for one, and pulled aggregate counts Georgia normally charges for. The state says it never got near the voter registration system or anyone’s individual records. Anthropic first mentioned the incident on 9 October without naming the state, and Fox Business named Georgia on Sunday. Source
Grok backed Republicans 92 per cent, GPT Democrats 97 per cent A studio led by Cornell’s Tapan Parikh asked Grok 4.7 and GPT-6.1 Sol who to vote for in nearly 500 midterm races and told them not to hedge, and they agreed less than one time in ten. With web search off, Grok refused to pick in most races while GPT still mostly went Democrat. The results came out on Saturday, and OpenAI says a developer test like this isn’t how people use ChatGPT. Source
Kimi’s new 3-trillion-class model goes open by 4 November Kimi K3.1 Pro is already live in Moonshot’s apps and API, and on Terminal-Bench, a test of coding agents, Moonshot puts it at 65.7 per cent against 66.4 for Claude Opus 5.5. Moonshot says it stripped hacking-related data out of training so the open version doesn’t push the frontier on cyberattacks. It went live on Friday. Source
Sweden’s government blocks ChatGPT and Claude on work computers from today Staff at the Government Offices are being sent to Microsoft Copilot instead, for security reasons, though some AI tools stay available through the government’s own internal services. It was announced on Friday, and Sweden’s Tax Agency already bans consumer AI tools. Source