Autonomous LLM Agents & CTFs: A Second Look

Collection of security-focused arXiv papers (May 22, 2026) covering multiple high-impact topics: (1) Critical crypto vulnerability on Polygon: systematic ECDSA nonce reuse by MEV searchers enables full private-key recovery via simple linear attacks; (2) LLM- and agent-related security: autonomous LLM agents for CTFs (general-purpose agents like claude-code match engineered ones but common failure classes remain) and THREAT, a potent jailbreak-prompt discovery framework that dramatically lowers safety-filter detection; (3) Adversarial ML: FRA-Attack improves transferability of image perturbions