Was I wrong about open models & cyber risks?

The rapid acceleration of critical vulnerability disclosures in October 2026, termed "Black October," serves as a catalyst for re-evaluating the actual risks posed by open-weight AI models versus closed systems. While critics previously warned that releasing powerful models like Claude Mythos would act as a catastrophic cyber weapon, the subsequent release of the highly capable GLM-5.3 has resulted in manageable industry turbulence rather than societal destabilization. Current evidence suggests that while AI models can successfully automate multi-stage exploits—such as a recent Firefox file exfiltration vulnerability identified by Epoch AI—the primary bottleneck for harm remains threat actor behavior and resource constraints rather than model access. The current surge in cyber incidents reflects an exponential trend line that predates modern AI, suggesting that the transition to an agent-mediated internet requires a broad societal shift in defensive infrastructure that transcends the debate between open and closed development silos.