Export controls for Fable are too late to slow proliferation (dualuse.dev)

Recent export controls placed on Anthropic's Fable model have sparked debate regarding their effectiveness in curbing the proliferation of advanced cybersecurity capabilities in AI. Although these controls aimed to mitigate systemic risks highlighted by recent cybersecurity breaches—such as account hijacking at major banks—many experts argue that they have come too late. Adversarial distillation has enabled others to transfer essential capabilities from frontier models into open-weight systems, making them even more adept at identifying vulnerabilities. Notably, Chinese models like GLM 5.2 have surpassed critical capability thresholds in cybersecurity, indicating that even if U.S. frontier models were entirely inaccessible, the potential for exploitation remains high. The implications are far-reaching: individuals and smaller entities have successfully leveraged open-weight models to replicate the capabilities of frontier systems, enabling them to find and exploit vulnerabilities at reduced costs. This unsettling trend underscores a shift in the landscape where traditional security measures may be rendered ineffective as open-access AI models proliferate. The call for stronger dual-use policies and proactive regulatory measures has gained urgency, as the AI/ML community faces the dual challenge of fostering innovation while ensuring responsible use—mirroring regulatory approaches in other high-stakes fields.