Claudini: AI autoresearch discovers SOTA jailbreak algorithms with 80% ASR
maksym_andr · x · 2026-10-21
The paper "Claudini" shows AI agents can discover novel adversarial attack algorithms for LLMs. Deploying Claude Code and Codex in an autoresearch loop with 30+ prior methods and a fixed compute budget, the pipeline jailbreaks GPT-OSS-Safeguard-20B at up to 80% ASR on CBRN queries (vs <50% for prior methods) and achieves 100% prompt-injection ASR against the adversarially robust Meta-SecAlign-70B (vs 82%). Notably, attacks developed on unrelated surrogate models generalize directly. A quote-tweet context lists related safety research from Tübingen, ETH Zurich, Cambridge, and Northeastern groups.
More from Safety
- Anthropic used Claude to autonomously shrink attack surfaces across 25 major OSS codebases — chrisrohlf · 2026-10-21
- Trusted intermediate CA 'leaks' with private key equal to 4, perfectly recreating the XKCD joke — arthurcolle · 2026-10-21
- Cloudflare discloses compromised eMudhra intermediate certificates, urges patching — Cloudflare Blog · 2026-10-21
- Package manager outage forces skipping release-age checks, ideal for supply chain attacks — sloppenheimer · 2026-10-21
- Same Text, Opposite Verdicts: DIY Test Exposes Unreliable AI Detectors — CommonExplanation711 · 2026-10-21
- How are you building your enterprise AI agent security stack for 2026? — Medical-Analyst2671 · 2026-10-21