Using LLMs to Exploit Vulnerabilities

A blog post highlights research demonstrating that teams of LLM agents (HPTSA) can coordinate to exploit real-world zero-day vulnerabilities by using a planning agent to spawn specialized subagents, improving success rates up to 4.5× across a benchmark of 15 vulnerabilities. While these agents are not discovering new flaws, they can autonomously explore and execute exploitation paths for previously unseen issues, signaling progress toward more fully autonomous AI-driven cyberattack capabilities.