Open models have become one of the defining fault lines in artificial intelligence. They are praised for democratizing access to cutting-edge capabilities, enabling independent researchers, smaller companies and public-interest groups to build on systems that would otherwise remain locked behind corporate APIs. But the same openness that makes these models attractive also makes them harder to contain, easier to repurpose and more difficult to govern.
Anthropic's latest warning has sharpened that debate. The company said a powerful new model released by Z.ai, a Chinese AI firm, could pose a major cyberattack risk. The concern is not simply that advanced models can answer technical questions; it is that sufficiently capable systems may lower the cost and complexity of malicious activity, including phishing, vulnerability discovery, malware assistance and other forms of cyber abuse. In other words, the risk is not theoretical. It is a question of how quickly frontier models can be adapted into tools that scale harm.
Openness Cuts Both Ways
The case for open models is compelling. They reduce dependence on a handful of large vendors, support reproducible research and allow organizations to fine-tune systems for specialized tasks without handing over sensitive data to external providers. In many markets, open models are also a strategic counterweight to the concentration of power in U.S. and Chinese tech giants. For governments, universities and startups, they can be the difference between participating in the AI economy and being locked out of it.
Yet the same properties that make open models useful also make them difficult to police. Once weights are released or a model is widely mirrored, control becomes diffuse. Safety filters can be stripped away, guardrails can be modified and the model can be deployed in environments far removed from the original developer's oversight. That creates a governance challenge unlike the one posed by closed systems, where access can be throttled, monitored or revoked.
Anthropic's warning arrives at a moment when policymakers are already struggling to define what responsible AI release should look like. The industry has long debated whether model capabilities should be disclosed gradually, whether certain weights should remain restricted and whether independent audits should be mandatory before release. The latest cyber-risk concerns are likely to intensify pressure for more formal standards, especially as models become more capable at code generation, reasoning and tool use.
Cyber Risk Meets Policy
The cyber dimension is especially sensitive because it is one of the clearest pathways from abstract model capability to real-world harm. A model that can help write code, explain systems and automate workflows can also be used to accelerate reconnaissance and exploit development. Even if a model does not directly create novel attack techniques, it can still reduce the skill threshold needed to execute them. That makes the issue broader than any single model or company.
For regulators, the challenge is to avoid blunt restrictions that could slow beneficial innovation while still addressing credible misuse risks. For developers, the task is to build systems that are useful without becoming easy instruments of abuse. For users, the lesson is that model performance alone is no longer the only metric that matters. Release strategy, access controls, monitoring and post-deployment safeguards are becoming central to the evaluation of frontier AI.
The Z.ai episode also reflects the increasingly global nature of the open-model ecosystem. Innovation is no longer confined to Silicon Valley. Chinese firms, U.S. labs and open-source communities around the world are all pushing the frontier, often with different assumptions about transparency, competition and state oversight. That global competition may accelerate progress, but it also makes common safety norms harder to establish.
The broader question is whether the industry can preserve the benefits of openness without normalizing the release of systems that materially expand cyber capability. That is not an easy balance. But as models grow more powerful, the cost of getting it wrong rises with each release.
Geothermal's Quiet Promise
The newsletter's focus on open-model trade-offs sits alongside another technology story with long-term significance: geothermal power. Unlike the AI debate, geothermal is not defined by speed or spectacle. Its appeal lies in reliability, low emissions and the possibility of firm clean power that can run around the clock. That makes it especially relevant at a time when data centers, electrification and industrial demand are putting new pressure on power systems.
Geothermal has long been viewed as promising but constrained by geography, drilling costs and project risk. Recent advances in drilling techniques, subsurface mapping and enhanced geothermal systems are helping to change that calculation. If those gains continue, geothermal could move from a niche resource to a more meaningful part of the clean-energy mix, particularly in regions where conventional renewables need firm backup.
Taken together, the two stories point to a common theme: technological progress is increasingly inseparable from governance. In AI, the question is how to release powerful models without amplifying cyber threats. In energy, the question is how to scale a dependable clean resource that has been technically feasible for years but commercially elusive. Both are reminders that the next phase of innovation will be judged not only by what it can do, but by how safely and sustainably it can be deployed.