AI Arms Race Makes Slowing Development a Risky Defense Strategy

The coordination problem with AI restraint

Calls to slow artificial intelligence development face a fundamental game-theory obstacle: restraint only works when everyone participates. Any organization or nation considering a pause must calculate what happens when competitors continue building more capable systems—and every other player faces the same dilemma.

Without credible mechanisms for global coordination and verification, AI capabilities will continue advancing regardless of individual decisions to slow down. This creates a strategic bind where developing powerful AI carries risks, but so does falling behind while adversaries forge ahead.

Why it matters

Autonomous AI systems are fundamentally changing cybersecurity economics by eliminating the time and expertise constraints that previously limited attackers. Organizations that assume they can rely on obscurity or that adversaries won't invest in complex attack chains are operating with outdated threat models.

How AI changes attack economics

Human attackers face real constraints. Reconnaissance, vulnerability analysis, testing, and chaining multiple weaknesses into viable exploits requires hours or days of skilled work. Many potential attack paths simply aren't worth the investment.

Autonomous systems eliminate these constraints. They can continuously perform reconnaissance, analyze vulnerabilities and permissions, test hypotheses, and search for ways to connect weaknesses—all at marginal costs approaching zero. Attack paths previously too difficult or time-consuming for humans suddenly become viable when machines can relentlessly explore what human analysts might overlook.

Rethinking defensive priorities

This shift forces defenders to ask different questions. Rather than simply assessing individual vulnerability severity, security teams must now determine which attack paths autonomous attackers can actually complete and where to disrupt them most effectively.

The answer isn't always patching. Effective controls might include revoking permissions, disabling dormant identities, rotating credentials, restricting access, or isolating systems. The objective shifts from fixing every weakness to disrupting meaningful attack paths and limiting blast radius when prevention fails.

Humans cannot continuously reason across every identity, permission, vulnerability, and system combination in large enterprises. When attackers operate at machine speed, defensive systems increasingly must as well.

Balancing capability and safety

This doesn't mean removing humans from security decisions or abandoning AI safety. People remain responsible for consequential decisions and must establish boundaries for autonomous system behavior. Guardrails, evaluations, governance, and responsible development become more critical as AI grows more powerful.

But these safeguards cannot be the only defense. Cybersecurity has always operated on a core principle: design defenses around what attackers can do, not what we hope they'll do.

The challenge isn't choosing between AI progress and AI safety—organizations need to advance capability, safety, and defense simultaneously. Slowing AI development may reduce some risks, but without assurance that adversaries will do the same, defenders must prepare for a world where increasingly capable AI exists and operates at speeds humans cannot match.

These arguments were made by Roie Cohen Duwek, co-founder and CTO at Surf AI, in an analysis first published on Calcalist.