In a demonstration of artificial intelligence capabilities, researchers have successfully infiltrated over half of their test websites using autonomous teams of GPT-4 bots. These bots, exhibiting remarkable coordination and the ability to spawn new bots as needed, exploited real-world ‘zero day’ vulnerabilities (flaws disclosed after the AI model’s training cutoff, so unknown to the agents).
Autonomous Exploits and the Evolution of AI
Just a few months ago, a research team made headlines by leveraging GPT-4 to autonomously exploit one-day (N-day) vulnerabilities—security flaws that are recognized but remain unpatched. Given each vulnerability’s Common Vulnerabilities and Exposures (CVE) description, GPT-4 could exploit 87% (13 of 15) of the one-day vulnerabilities in the researchers’ test set; without the description, it managed 7%. This breakthrough was significant, showcasing the potential of AI in identifying and manipulating known vulnerabilities.
Fast forward to this week, and the same group of researchers has released a follow-up study. They claim to have successfully exploited zero-day vulnerabilities—flaws unknown to the agents ahead of time—using a team of autonomous Large Language Model (LLM) agents in which a planning agent launches specialized subagents. This was achieved through a method known as Hierarchical Planning with Task-Specific Agents (HPTSA).
HPTSA: A New Paradigm in AI Coordination
HPTSA represents a significant shift in how AI systems tackle complex tasks. Rather than relying on a single LLM agent to manage numerous intricate tasks, HPTSA employs a “planning agent” that orchestrates the entire process. This planning agent deploys multiple task-specific “subagents,” each an expert in its designated function. This hierarchical approach mimics a traditional business structure where a manager delegates tasks to specialized subordinates, thereby alleviating the burden on any single agent.
AI Teamwork
When benchmarked against 15 real-world web-focused vulnerabilities, HPTSA was 2.7 times as successful as a solitary LLM agent over five attempts, and up to 4.5 times as successful on a single attempt. The autonomous team successfully hacked 8 out of 15 zero-day vulnerabilities, while a single LLM managed to infiltrate only 3 of the 15 vulnerabilities.
Correction, 26 October 2026: this article said GPT-4, given the CVE list, could exploit 87% of critical-severity CVEs; the researchers gave GPT-4 each vulnerability’s CVE description and it exploited 13 of a test set of 15 one-day vulnerabilities, only some of them critical; it also said HPTSA was 550% more efficient than a single agent, but the paper reports up to 4.5 times on a single attempt and 2.7 times over five attempts; it also called the agents self-replicating, which the paper does not.