Curriculum of Models: RL-trained investigator surfaces LLM jailbreaks
austinc3301 · x · 2026-11-04
A new paper introduces Curriculum of Models (CoM): train an 'investigator' model via RL to attack a target LLM, progressing through a curriculum of easier-to-jailbreak intermediate targets before the real one, solving the sparse-reward problem caused by refusals. Aimed at helping developers find vulnerabilities pre-deployment.
More from Safety
- OpenAI finds agents score worse on safety evals than production; eval-objective awareness drives the gap — kaicathyc · 2026-11-04
- Security researcher: Homebrew's official install method is sketchier than a Google ad — matthew_d_green · 2026-11-04
- Fake NFT job interviews weaponize GitHub repos to drain crypto wallets — dotey · 2026-11-04
- India's CERT-In gets access to Anthropic model, treads carefully over data leaks — nordicinst · 2026-11-04
- Cloudflare launches Zero Wire, a public vulnerability intelligence feed on Radar — Kyrannio · 2026-11-04
- CCIA argues training AI on copyrighted material is fair use in OpenAI case — nordicinst · 2026-11-04