Import AI #453 covers three major AI developments: METR and Epoch's MirrorCode benchmark showing Claude Opus 4.6 can autonomously reimplement complex software (e.g., a 16,000-line bioinformatics toolkit), Google DeepMind's taxonomy of six attack vectors against AI agents, and the Windfall Policy Atlas organizing 48 policy proposals for transformative AI governance.
Safety
Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment
Claude Opus 4.6 autonomously reimplements 16,000-line bioinformatics tools in the MirrorCode benchmark, while researchers catalog attack vectors against agents and policymakers organize 48 governance proposals for transformative AI.
Monday, April 13, 2026 12:00 PM UTC2 MIN READSOURCE: Import AI (Jack Clark)BY sys://pipeline
Tags
safety
/// RELATED
ResearchApr 24
Machine Learning Reveals Unknown Transient Phenomena in Historic Images
ML analysis of 108k historic transients finds statistical correlation with nuclear testing windows, validating disputed optical phenomena as real astronomy rather than plate defects.
ProductsApr 28
DJI’s new Mic Mini 2 adds colorful covers to help them blend in
DJI launches the Mic Mini 2 with 12-color swappable magnetic covers at €59, betting on user customization and aesthetic blending over technical spec bumps.