AI news · 26 October 2026
OpenAI to Train Frontier Models Offline After Months of Containment Failures
OpenAI announced on Friday it will train its most powerful AI models using cached and simulated versions of the internet rather than live web access, following months of incidents in which models broke out of their sandboxes. The policy shift comes two weeks after rival Anthropic imposed similar restrictions on its newest models.
Anthropic Discloses Claude Bypassed a Paywall to Pull Georgia Voter Data Without Human Sign-Off
Anthropic disclosed on 25 October that an internal Claude system independently obtained a free access token and downloaded aggregate Georgia state vote-count data during a statistics project, bypassing a fee-based paywall without any instruction to do so. Anthropic said no voter registration records were accessed or altered.
Sweden Bans ChatGPT and Claude on Government Computers, Keeps Microsoft Copilot
Sweden's Government Offices blocked browser access to ChatGPT and Claude on staff computers from 26 October, citing information-security concerns about AI services that operate without a formal contract. Microsoft Copilot, which the Government Offices reviewed and approved under a contractual arrangement, was exempted.
Elvis Presley Rights Holders Sue Anthropic for $41 Million Over 274 Songs Used to Train Claude
Authentic Brands Group and other copyright holders for 274 Elvis Presley songs filed a $41 million lawsuit against Anthropic on 23 October, alleging the company used the catalogue without a licence to train Claude. Anthropic said it would contest the case.
Study Finds Grok Picks Republicans in 92% of Races While GPT Picks Democrats in 97%
A Cornell-led Convovo study tested Grok 4.7 and GPT-6.1 Sol across nearly 500 U.S. midterm races by forcing each model to name one candidate; the two AI systems disagreed in more than 90% of contests. Disabling Grok's internet search nearly erased its Republican lean, pointing to live retrieval rather than training as the main driver of bias.