Import AI 447: The AGI economy; testing AIs with generated games; and agent ecologies
Two items - LLM-enabled bioweapon skill uplift and agentic AI security failures - surface concrete risk themes directly relevant to APS AI governance and biosecurity policy.
Key points
- Import AI 447 covers AGI economics, bioweapon uplift from LLMs, AI agent security failures, and robotics deployments.
- The agent ecology study and bioweapon uplift research carry the most direct relevance for APS AI governance and risk practitioners.
- This is a curated research newsletter; individual papers warrant separate engagement for deeper analysis.
Implications for Australian agencies
- Monitor AI governance and biosecurity policy teams may want to monitor the LLM bioweapon uplift research as evidence that frontier models materially lower barriers to dual-use biological knowledge.
- Consider Agencies evaluating or deploying agentic AI systems could consider whether their current risk frameworks address the failure modes documented in the Agents of Chaos study, including prompt injection, non-owner compliance, and resource looping.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 2 March 2026
"Import AI 447: The AGI economy; testing AIs with generated games; and agent ecologies"
Source: Import AI – Substack (Jack Clark)
Published: 2 March 2026
URL: https://importai.substack.com/p/import-ai-447-the-agi-economy-testing
Import AI issue 447 covers five distinct research and industry developments. A MIT/WashU/UCLA paper models an AGI-driven economy where human value shifts to verification and oversight of machine agents. A multi-institution study finds LLMs increase novice accuracy on bioweapon-related tasks by roughly 4x, raising dual-use concerns. An AI GAMESTORE benchmark shows frontier LLMs perform below 30% of human baseline on simple web games. Physical Intelligence reports early commercial robot deployments using vision-language-action models. An 'Agents of Chaos' study exposes significant security and reliability failures in deployed AI agents, including prompt injection, resource looping, and non-owner compliance - findings relevant to any agency considering agentic AI deployments.
Implications for Australian agencies:
- [Monitor] AI governance and biosecurity policy teams may want to monitor the LLM bioweapon uplift research as evidence that frontier models materially lower barriers to dual-use biological knowledge.
- [Consider] Agencies evaluating or deploying agentic AI systems could consider whether their current risk frameworks address the failure modes documented in the Agents of Chaos study, including prompt injection, non-owner compliance, and resource looping.
Retrieved from SIMS, 18 July 2026.