AI Scales Cyber Offense, Boosts Startups 1.9x Revenue
Frontier models hit 50% success on expert-level cyber tasks taking 3h; AI-adopting startups gain 44% more use cases, 1.9x revenue, 39% less capital need; automation rises gradually to 90% success on hours-long tasks by 2029.
Frontier AI Scales Cyberattack Capabilities Predictably
AI systems follow scaling laws in offensive cybersecurity, with capabilities doubling every 9.8 months since 2019 (5.7 months since 2024). Latest models like GPT-5.3 Codex and Opus 4.6 achieve 50% success on tasks requiring human experts 3.1-3.2 hours, using benchmarks including CyBashBench, NL2Bash, InterCode CTF, NYUCTF, CyBench, CVEBench, CyberGym, plus a new 291-task dataset calibrated by 10 pros. Evaluated models span GPT-2 (2019) to GLM-5 (2026 open-weight, lagging closed-source by 5.7 months). This dual-use risk means defensive tools easily flip to offense, amplifying policy challenges as AI becomes an 'everything machine'—strong in biology also aids bioweapons, code vuln finding aids hacking.
Deep AI Adoption Accelerates Startup Performance
Startups trained on real AI integrations discover 44% more use cases (2.7 additional), focused on product dev and strategy, yielding 12% more tasks completed, 18% higher odds of paying customers, and 1.9x revenue versus controls. Each extra use case adds 0.85 tasks and 26% revenue. Examples: Gamma auto-generates product variants from usage patterns (1 PM ships team-scale features); Ryz Labs parallel-AI-codes PRDs for diverse prototypes; FazeShift AI-skips human accounts receivable; Ranger bootstraps traction pre-funding for better terms. In INSEAD's AI Founder Sprint (515 high-growth firms, $25k credits/models), treated firms cut capital needs 39.5% ($220k less) without labor hikes, enabling capital-efficient scaling like past tech waves (e.g., internet-era Amazon). Bottleneck is managerial mapping, not tech—education unlocks value.
Automation Rises Gradually, Forecasts Mute GDP Impact
MIT analysis of 3k O-NET tasks (17k worker evals) shows 'rising tides' over 'crashing waves': frontier models shift 50% success from 3-4h to 1-week tasks (2024Q2-2025Q3), flat success-vs-duration slope across job families. By 2029, 80-95% success projected for most text/partially-text tasks (90% for hours-long), validating METR's time horizons. Yet Forecasting Research Institute's survey (69 economists, 52 AI/policy experts, 38 superforecasters, 401 public; Oct2025-Feb2026) predicts moderate-rapid AI progress by 2030 (basic-to-high research/creativity/physical tasks) but only ~1pp GDP growth boost (from 2.4%), stable TFP/labor participation, rising inequality. Economists favor retraining/unemployment insurance/AI Manhattan Project over UBI/compute tax/job guarantees; 14% chance short-term GDP/inequality surge. Paradox: experts expect capability jumps but dampened macro effects, clashing with lab visions of explosive change.