Dreadnode (arxiv, May 2025) presents an agentic AI red-teaming system that autonomously orchestrates adversarial attacks, 45+ strategies, 450+ transforms, 130+ scorers, via natural-language objectives, eliminating manual workflow construction. A case study against Meta's Llama Scout achieved approximately 85% attack success rate with critical-severity findings across ransomware generation, SQL injection, phishing, and self-harm content, using zero human-written code. The system exposes that current frontier safety training remains systematically bypassable at scale through automated multi-step attack chains. For financial institutions deploying or evaluating agentic AI systems, this signals that adversarial robustness cannot be assumed from safety certifications alone and that the cost of adversarial probing is falling rapidly, raising both the bar for vendor due diligence and the threat surface for internal deployments.
313 other tracked developments touching the same claims.