A February 2026 preprint from researchers at Monash, UCLA, UCSB, and NUS argues that AI agents break the historical labor-scarcity assumption underpinning cybersecurity economics: because agents can probe thousands of targets at near-zero marginal cost, even low per-attempt success rates become profitable, extending viable attacks to previously ignored "long-tail" systems. The paper contends that existing model-level safeguards (data filtering, alignment, guardrails, access controls) are structurally insufficient against adversaries who control open-weight models, and proposes that defenders must build frontier offensive agents operating in audited cyber ranges. For financial institutions, this structural shift in attack cost curves implies that cyber risk underwriting assumptions and enterprise security vendor ROI models built on attacker resource constraints may require revision as agentic capabilities mature.
345 other tracked developments touching the same claims.