The latest edition of The Download underscores a central tension shaping the frontier AI market: open models can broaden access to advanced capabilities, but that same accessibility may also widen the attack surface for malicious actors. Anthropic's recent report on a new model released by Z.ai, a Chinese artificial intelligence company, has added urgency to that debate by arguing that the system could pose a major cyberattack risk if used irresponsibly.
Open Models, Real Gains
Open models have become one of the most consequential forces in artificial intelligence because they allow developers, researchers, and startups to inspect, adapt, and deploy powerful systems without relying entirely on closed commercial platforms. For many in the industry, that openness is not just a philosophical preference; it is a practical advantage. It can reduce costs, speed experimentation, and help smaller players compete with dominant AI labs that control proprietary systems.
The appeal is especially strong in sectors where customization matters. Enterprises may want to fine-tune models for internal workflows, governments may seek greater transparency, and academic researchers often need access to model weights to test safety, bias, and robustness. In that sense, open models can democratize AI development and accelerate the diffusion of useful tools across the economy.
Yet the same features that make open models attractive also make them harder to contain. Once a capable system is released broadly, developers lose much of the control they would have over a closed product. That creates a structural dilemma: the more powerful the model, the more valuable openness becomes for legitimate users, and the more dangerous it may become if repurposed for harmful ends.
Cyber Risk Comes Into Focus
Anthropic's warning about Z.ai's model places cybersecurity at the center of the open-model debate. The company's concern is not simply that a model can generate code or answer technical questions, but that a sufficiently capable system could help lower the skill threshold for cyber intrusions. In practical terms, that could mean assisting with reconnaissance, vulnerability discovery, phishing content, or other steps that make attacks easier to execute at scale.
This is a significant shift in how the AI industry is discussing risk. Earlier debates often focused on misinformation, copyright, or labor displacement. Those remain important, but cyber risk is increasingly viewed as a more immediate and measurable threat because it can translate directly into financial damage, infrastructure disruption, and national security concerns.
The issue is particularly sensitive when models are released by companies operating in different regulatory and geopolitical environments. Open-source or open-weight releases can spread globally within hours, making it difficult for any single company or government to monitor downstream use. That reality complicates efforts to establish safety standards, because even a well-intentioned release can be repurposed in ways the original developer did not anticipate.
Anthropic's intervention also reflects a broader competitive dynamic in frontier AI. Companies are under pressure to innovate quickly, but they are also being asked to prove that their systems are safe enough for wide release. In that environment, warnings about cyber misuse can serve both as genuine risk assessments and as signals in an increasingly intense race for credibility, influence, and market share.
Geothermal's Quiet Promise
The newsletter's broader framing also points to another frontier technology with outsized implications: geothermal power. While less visible than AI, geothermal energy is attracting attention as a potentially reliable source of low-carbon electricity for a world that needs more power for data centers, industrial electrification, and the AI boom itself.
That connection matters. As AI models grow larger and more energy-intensive, the question of where that electricity comes from is becoming inseparable from the future of the sector. Geothermal power offers a compelling proposition because it can provide steady, round-the-clock generation rather than intermittent output tied to weather conditions. For AI infrastructure, that kind of baseload power is especially valuable.
The promise of geothermal is not without constraints. Development can be capital-intensive, site-specific, and slow to scale. But in a technology landscape increasingly shaped by energy demand, it stands out as one of the few clean power sources that could support the long-duration, high-load needs of advanced computing.
Taken together, the two themes in this edition of The Download point to a broader reality: the next phase of technological progress will be judged not only by what systems can do, but by how safely they can be deployed and what infrastructure is required to sustain them. Open models may expand access to AI, but the cost of that openness is rising. And as the industry pushes deeper into frontier capabilities, energy and security are becoming as central to the story as the models themselves.