AI agent safety research by Kody Abbott. Learning in public.
Each folder is a research project. notes.md is the trail, README.md is the report.
Research is AI-assisted using Claude Code (Anthropic; currently Claude Fable 5, earlier projects Opus 4.6). I direct and verify, the agent searches and drafts. Project READMEs are marked with an AI-ASSISTED-NOTE banner to make this transparent. See CLAUDE.md for what that means and how this repo works.
Repo structure based on Simon Willison's simonw/research. Simon's work on the Lethal Trifecta and prompt injection is a primary source here.
The .claude/ directory contains the agent team that runs this research:
agents/-- specialized agents (researcher, source-verifier, adversarial-reviewer, url-auditor)rules/-- research workflow and source verification rules, auto-loaded every sessionsettings.json-- project permissions and model config
See CLAUDE.md for full agent instructions.
- jensen-gtc-agent-safety -- Jensen Huang's GTC comments on AI agent capability constraints
- simon-willison-lenny-podcast -- Simon Willison on Lenny's Podcast: AI state of the union, lethal trifecta, and agent security
- local-model-benchmarks -- Same-day benchmarks of trending open models on one known machine, with methodology stated so the numbers can be argued with
- anthropic-only-company-x-thread -- The Aug 2026 Baker/Douglas/Amodei X exchange on Anthropic, regulation, and open weights: curated full-text report of the public figures, IDs-only dataset (X Developer Policy), and collection cost analysis (X API vs browser automation)