The Cyber Risk Discourse is Broken

The Cyber‑Risk Discourse Is Broken

The conversation around AI‑driven cyber threats has hardened into two opposing camps: one that cries for a blanket ban on open‑weight models, and another that argues that only closed, proprietary systems can keep society safe. Both positions overlook the complex trade‑offs, the realities of global AI development, and the fact that the same tools can be used for defense as well as offense. A more nuanced discussion is needed—one that asks the right questions rather than settling for slogans.

1. Why the Binary Framing Fails

• Open‑weight ≠ Open‑danger, Closed ≠ Closed‑safe – Evidence shows that many documented cyber incidents have involved closed‑model APIs, while open‑weight releases have not yet produced a comparable wave of attacks. The assumption that openness automatically equals risk, and that secrecy guarantees safety, is therefore misleading.
• Policy decisions are interdependent – Banning open models while allowing closed models to proliferate creates an asymmetry: defenders lose a valuable, auditable tool, while attackers continue to exploit the more powerful, less‑controlled closed systems.
• Trade‑offs matter – Even if open models are marginally easier to misuse, they also enable rapid defensive innovation, especially on air‑gapped or highly regulated networks where closed APIs are impractical.

2. The Missing Global Perspective

• China’s risk posture is often caricatured – Chinese AI firms must register major model releases with the state, and their risk assessments are shaped by a different set of political and social incentives. Dismissing them as “unsafe by default” ignores the fact that their oversight mechanisms, however opaque, differ fundamentally from the U.S. approach.
• Different incentives, similar challenges – Both Western and Chinese labs grapple with the cost of safety testing. Comprehensive evaluations can run into tens of millions of dollars in compute, and the decision of how much to invest is a business judgment as much as a technical one.
• Cross‑border learning is scarce – Public discourse rarely engages with how Chinese companies balance safety and competitiveness, leading to ad‑hominem arguments rather than evidence‑based analysis.

3. Rethinking the Policy Puzzle

1. Define clear, measurable safety thresholds – Instead of calling for outright bans, policymakers should ask: What is the minimum compute and testing a lab must perform before releasing a model? Setting transparent benchmarks can level the playing field.
2. Promote responsible diffusion – Encourage the development of open‑weight models that are auditable, can be run on private infrastructure, and come with community‑driven guardrails. This reduces reliance on a duopoly of closed‑model providers.
3. Invest in defensive tooling – Allocate resources for security teams to adapt open models for detection, attribution, and mitigation of AI‑enhanced attacks. The defensive value of open models often outweighs their potential misuse in the short term.
4. Create a balanced regulatory framework – Regulations should address both open and closed models, ensuring that APIs are not exempt from safety requirements while also protecting legitimate research and innovation.

4. Embracing Nuance Over Fear‑mongering

The most vocal anti‑open‑weight arguments predict a sudden, catastrophic escalation of cyber harm if models like GLM‑5.3 remain public. So far, the data does not support that claim. Instead, the evidence points to a gradual, uneven diffusion of capabilities across both open and closed ecosystems. Recognizing this reality allows us to:

• Focus on mitigation rather than prohibition, building safeguards that evolve with the technology.
• Foster international dialogue, learning from differing governance models rather than demonizing them.
• Avoid a lose‑lose outcome where the U.S. hampers its own competitiveness while inadvertently widening the offense‑defense gap.

---

Closing Thoughts

The cyber‑risk discourse is broken because it treats a multifaceted problem as a simple choice between “open equals danger” and “closed equals safety.” Effective policy must move beyond this false dichotomy, incorporate global perspectives, and prioritize concrete, measurable safeguards for all AI models. By doing so, we can protect critical infrastructure without stifling the defensive innovations that open‑weight models uniquely enable.