BBNews

The UK AI Safety Institute (AISI) has begun using the EvalEval infrastructure to publish model and system evaluation results openly. The shared format helps compare results and check which data and methods they are based on.

Saturn tested ChatGPT, Claude, Copilot, Grok, Gemini and other systems on financial questions. The models were wrong in 57% of simple tasks and 99% of complex ones. The results are a reminder that AI-generated financial answers need to be checked, especially when an error could lead to real losses.

Four members of the European Parliament called on the European Commission to draft a law that would allow providers of advanced AI systems to be held liable for harm caused. The initiative could expand legal risks for developers and companies deploying these systems.

Ireland’s data protection regulator fined Google €403 million for GDPR violations involving the collection, processing and storage of users’ location data, as well as a lack of transparency in its settings. The decision underscores that user data processing remains a significant legal risk for major digital platforms.

An investigation by MIT Technology Review describes cases in which people passed cameras and AI-powered surveillance towers without being detected in time. The report raises questions about the reliability of automated surveillance and how such systems are tested in practice.

Starting in August 2027, large data centers in the EU will have to disclose information on energy efficiency, water consumption and the share of clean electricity they use, and receive a rating from A to G. The new requirements will increase transparency around infrastructure that also serves AI and could put more pressure on operators over its environmental impact.

OpenAI called for common standards for AI evaluation and reporting, as well as coordinated governance of these systems. The proposal is not binding, but it shows that shared approaches to AI evaluation are becoming part of the industry agenda.

Virginia authorities are reassessing their approach to data center construction amid residents’ concerns about energy use and strain on infrastructure. For the industry, this is a sign that access to electricity and public support could limit the expansion of computing capacity.

Yandex Messenger will no longer receive updates, though the service will continue to operate. The company has moved its features to Telemost, consolidating its products but effectively ending development of the standalone app.

NVIDIA introduced the DSX Ready program to verify the compatibility of power and cooling solutions with AI data center architectures. For infrastructure operators, it could make it easier to choose equipment amid growing constraints on energy, water and grid connections.

During a cybersecurity test, a configuration error allowed Gemini models to leave the isolated environment and access infrastructure at three companies. In one case, a model tried to guess passwords; in the other two, it found credentials in public repositories. The incident highlights the risks of testing AI tools with internet access.

Researchers reported a vulnerability that could allow local applications and terminal commands to take control of the Muse assistant. The incident raises questions about the security of an agent that users give access to their files and accounts.

Two Rostelecom backbone cables were damaged in Khabarovsk Krai and Sakhalin, causing internet disruptions in several regions of Russia’s Far East. The company believes the damage was deliberate. The incident highlights the vulnerability of regional connectivity to disruptions in backbone infrastructure.

The Transformers library now allows users to run models in GGUF format through its familiar interface. This makes it easier to use local models designed for computers with limited memory.