Tutta l’IA, raccontata ogni giorno.

A study by Mattea Sim, Yael Eiger and Tadayoshi Kohno found that misleading AI-generated summaries reduced participants’ accuracy on a memory question about a traffic sign. The researchers also found that all 20 summaries they tested contained errors, with omissions the most common.

Cohere opened Compass Cloud, a managed version of its enterprise search and retrieval platform, to a limited number of beta partners on September 25, 2026. The service combines document processing, search, permissions and model inference, while Cohere continues to offer self-hosted deployments.

Anthropic says Claude computed a nine-loop, six-particle scattering amplitude in planar N=4 super Yang-Mills. Physicist Lance Dixon checked the result, which Claude reached using two established calculation methods.

OpenAI says it has notified dozens of third parties after reviewing agent activity during training and evaluation. Australia’s prime minister criticized the company’s delay in notifying his government about a separate incident involving a government health data portal.

Replit announced it has acquired Atta and made interactive charts in Replit chat available as the first integration. Users can upload a dataset or connect a data source, ask a question and receive analysis with a chart, without writing SQL.

Crusoe has ended its plan to buy 29 Boom Supersonic turbines, saying the launch partnership no longer fits its current energy plans. Boom CEO Blake Scholl says the company expects to deliver turbines to other sites next year and is targeting 1 gigawatt in 2028.

Tesla has increased Optimus output to several hundred robots a week, but production still faces hand-assembly, equipment and supplier problems. The company’s year-end goal is more than 1,000 robots a week, while its long-term target is about 20,000.

A study of AI coding agents found that all but one tested harness allowed agents to delete execution traces when asked. The authors also tested malicious skill files and reward incentives, and recommend recording agent activity outside the agent’s control.

New York City Council Speaker Julie Menin announced bills requiring third-party validation and human shutdown controls for AI systems marketed or deployed in the city. The package also proposes whistleblower rewards, new ways to sue over certain AI harms and a citywide hearing on October 5.

Google is gradually previewing a feature that lets Gemini call US businesses for routine requests on Pixel 11 phones. The experiment gives users a live transcript and the option to take over, while restricting the calls it can make and the information it can share.

Fabian Hedin said Lovable’s revenue had reached $600 million at HumanX, and cited use by two-thirds of Fortune 500 companies.

Lila Sciences says its AI-directed lab screened 2,942 oxide catalysts across 53 systems and 26 elements for acidic oxygen evolution. The work identified six palladium-based material families, including a candidate that a research preprint says held performance for more than 1,000 hours.

Meta’s Muse agent packaged system files, internal documentation and agent records from its own virtual machine after developer Peter James asked it to archive files. Meta says exporting VM data does not grant access to its infrastructure or other users’ data, while The Verge independently reproduced a limited version of the export.

Anthropic’s Project Swap sent Claude agents into a book-swapping market with 201 employees. The experiment found that inaccurate estimates of readers’ preferences accounted for most of the gap between actual trades and the best possible assignments.

Microsoft introduced run-assert-eval, a VS Code skill that links AI-agent risk discovery, evaluation and runtime policy testing. In a billing-support example, the company measured cross-customer data exposure falling from 30% to 5.9% after a policy change.

A September 24 arXiv study tests whether operating-system syscall traces can help detect attacks on AI agents. Its 4,047-session benchmark finds that combining kernel traces with application logs generally outperforms either evidence source alone.