The Download explores the trade-offs of open models and the promise of geothermal power.

Anthropic has published a report arguing that a powerful new model from Chinese company Z.ai poses a major cyberattack risk, primarily because it can be freely downloaded and modified by anyone. The warning has reignited a broader argument over whether frontier AI models should be released openly or kept inside the labs that build them.

Safety advocates say open models can be misused by bad actors. Supporters counter that openness prevents a small number of companies from controlling a transformative technology. The dispute is more complicated than either camp tends to admit, as MIT Technology Review explored in a recent analysis.

Background: How the open-model fight took shape

The tension is not new. For years, labs have split into two camps: those that publish model weights for anyone to inspect and run, and those that restrict access through APIs and gated releases. Anthropic's report on Z.ai's model is the latest flashpoint, arriving as capable open systems close the gap with closed ones.

Development: Meta's new models and Anthropic's replication warning

Meta has launched a set of new AI models built for complex, multi-step tasks, aiming to compete with OpenAI and Anthropic. The models scored well on tests of reasoning, writing, and coding, and CEO Mark Zuckerberg said they bring Meta to the "frontier" of AI. They are expected to soon power the Muse agent.

Anthropic separately warned that AI models may soon be able to self-replicate across devices without needing access to data centers, potentially spreading through smart TVs and home appliances.

What else happened this week

A cyberattack overheated fridges at hundreds of French supermarkets, pushing temperatures to 120°F, spoiling food, and sparking fires. The incident has raised alarm about cyber threats to food security and adds to a recent wave of cyberattacks and cybercrimes that AI is exacerbating.

Fervo Energy's Utah facility, the world's first utility-scale enhanced geothermal plant, is now online and supplying carbon-free electricity. The company uses horizontal drilling and hydraulic fracturing to open pathways through underground rock, and is betting AI's power boom will help scale enhanced geothermal. It has a $2.2 billion public offering and an expanded agreement with Google behind plans to build larger plants and cut costs. Whether the approach works beyond Utah and Nevada remains an open question.

Police have a new AI tool that identifies homes from interior photos, matching single pictures to Airbnb and Zillow listings. Police-tech giant Flock is tightening its rules.

Ukrainian President Zelenskyy claims drone strikes on Yandex data centres quietened Kremlin-funded social media accounts, though troll farm experts are unconvinced.

NASA's Perseverance rover has set a Mars driving record, covering 28 miles since landing in 2021 and passing Opportunity after detouring to study ancient rocks. The Solar Orbiter has traced how a mysterious solar switchback formed, a finding that could improve space weather forecasts.

Elon Musk became the 2026 midterms' biggest known donor, giving $158 million to a Republican Senate super PAC. Anduril and Palantir founders Palmer Luckey and Joe Lonsdale have backed execution livestreams, supporting Pete Hegseth's plan.

What to watch

The open-versus-closed model debate now turns on whether voluntary warnings and lab-level caution can meaningfully shape how freely weights circulate. Anthropic's replication warning and its Z.ai report both point in the same direction: the argument is shifting from philosophy to concrete risk claims.