Ex-Meta Expert: AI Security Policy Must Shift from Model Launches to Attacker-Defender Ecosystem

joshua_saxe · x · 2026-08-13

Joshua Saxe, former Chief Scientist at Meta, argues that current AI security policies are fundamentally flawed because they focus excessively on evaluating and blocking "model launches" for dual-use cyber risks.

He emphasizes that AI models themselves are not the direct cause of cyber harms; the real risk lies in the imbalance of the broader attacker-defender ecosystem. As cheap AI security agents become widespread, attackers will automate kill chains, forcing defenders to rely heavily on AI as well. Saxe recommends that policymakers shift their focus away from singular model control and instead measure and shape the macro-level "attacker-defender ecosystem" to navigate the upcoming era of cybersecurity instability.

Original post →

More from Safety

Safety channel →