Investigating the Potential Use of Frontier AI Models for Offensive Cyberattacks: A Human Uplift Study

Why It Matters

The findings temper fears of an immediate AI‑driven surge in sophisticated cyber attacks while highlighting a potential skill‑levelling effect that could expand the pool of capable low‑skill threat actors. Policymakers and security teams must balance AI guardrails with proactive defenses against faster, AI‑assisted reconnaissance.

Key Takeaways

• AI tools gave novices 18% higher success on easiest attack chain
• Overall AI uplift for full attack chains was statistically insignificant
• AI access cut novices' task time by 2.2×, technical participants 1.4×
• Model guardrails caused refusals; 40% needed three or more follow‑ups
• Study suggests AI may level skills but not guarantee complex exploits

Pulse Analysis

Frontier artificial intelligence models have leapt from simple text generators to systems capable of nuanced reasoning, prompting security experts to question how quickly these tools could be weaponized. While public discourse often dramatizes AI as a catalyst for next‑generation cybercrime, the technology’s actual impact depends on the skill set of the operator and the robustness of model safeguards. Early incidents of AI‑aided vulnerability scanning and exploit drafting have surfaced, but the broader threat landscape remains uncertain, especially for actors lacking deep technical expertise.

The RAND study, spanning September 2025 to January 2026, offers empirical insight into this uncertainty. By randomising 157 participants into AI‑assisted and control groups, the researchers measured performance across three attack chains of increasing complexity. Novice participants showed a measurable uplift—an 18‑percentage‑point jump in success on the easiest chain and a 2.2‑fold speed increase—suggesting AI can lower entry barriers for basic reconnaissance and exploitation. However, the same uplift vanished for more sophisticated chains, indicating that advanced offensive operations still demand expertise beyond what current models can provide.

These results carry weight for both regulators and enterprise security teams. The frequent refusals triggered by model guardrails underscore the importance of built‑in safety mechanisms, yet they also reveal a friction point that adversaries may learn to circumvent. As AI providers iterate on capability and policy, the security community must develop detection tools that account for faster, AI‑augmented attack phases while monitoring for emerging tactics. Continued interdisciplinary research will be essential to gauge whether future model improvements will shift the balance toward broader, more dangerous cyber threats.