Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over gDP forecasting
Rapidly scaling AI cyber-offensive capability challenges APS agencies' assumptions about threat timelines and defensive posture.
Key points
- Frontier AI models double in offensive cyber capability every 5.7 months, reaching 50% success on expert-level hacking tasks.
- Open-weight models lag closed-source frontier by only 5.7 months, meaning advanced cyber capabilities diffuse quickly into public access.
- A separate study found AI-adopting startups generated 1.9x more revenue - relevant context for APS workforce and productivity thinking.
Summary
Lyptus Research has documented clear scaling laws in AI offensive cybersecurity capability, with frontier models doubling in effectiveness roughly every 5.7 months and now achieving 50% success on tasks requiring 3+ hours of expert human effort. Open-weight models trail the closed-source frontier by less than six months, suggesting rapid proliferation of advanced offensive tools. A separate INSEAD/Harvard field experiment across 515 startups found that structured AI adoption education produced significant performance gains - 44% more AI use cases discovered, 1.9x higher revenues, and 39.5% lower capital demand - with authors noting that the bottleneck is managerial discovery, not technology access.
Implications for Australian agencies
- Monitor Cyber and ICT security teams across APS agencies may want to monitor Lyptus Research's ongoing benchmarking as it provides empirical evidence for updating AI-related threat assessments.
- Consider Policy and strategy teams could consider whether AI capability uplift programs for APS staff draw on the INSEAD finding that structured use-case education - not just tool access - drives meaningful productivity gains.
Implications are AI-generated. Starting points, not advice.
"Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over gDP forecasting" Source: Import AI – Substack (Jack Clark) Published: 6 April 2026 URL: https://importai.substack.com/p/import-ai-452-scaling-laws-for-cyberwar Lyptus Research has documented clear scaling laws in AI offensive cybersecurity capability, with frontier models doubling in effectiveness roughly every 5.7 months and now achieving 50% success on tasks requiring 3+ hours of expert human effort. Open-weight models trail the closed-source frontier by less than six months, suggesting rapid proliferation of advanced offensive tools. A separate INSEAD/Harvard field experiment across 515 startups found that structured AI adoption education produced significant performance gains - 44% more AI use cases discovered, 1.9x higher revenues, and 39.5% lower capital demand - with authors noting that the bottleneck is managerial discovery, not technology access. Implications for Australian agencies: - [Monitor] Cyber and ICT security teams across APS agencies may want to monitor Lyptus Research's ongoing benchmarking as it provides empirical evidence for updating AI-related threat assessments. - [Consider] Policy and strategy teams could consider whether AI capability uplift programs for APS staff draw on the INSEAD finding that structured use-case education - not just tool access - drives meaningful productivity gains. Retrieved from SIMS, 18 May 2026.