- Researchers recreated the Equifax hack and watched AI do everything without direct control
- The AI model successfully carried out a major breach with zero human input
- Shell commands weren’t needed, the AI acted as the planner and delegated everything else
Large language models (LLMs) have long been considered useful tools in areas like data analysis, content generation, and code assistance.
However, a new study from Carnegie Mellon University, conducted in collaboration with Anthropic, has raised difficult questions about their role in cybersecurity.
The study showed that under the right conditions, LLMs can plan and carry out complex cyberattacks without human guidance, suggesting a shift from mere assistance to full autonomy in digital intrusion.
You may like
From puzzles to enterprise environments
Earlier experiments with AI in cybersecurity were mostly limited to “capture-the-flag” scenarios, simplified challenges used for training.
The Carnegie Mellon team, led by PhD candidate Brian Singer, went further by giving LLMs structured guidance and integrating them into a hierarchy of agents.
With these settings, they were able to test the models in more realistic network setups.
In one case, they recreated the same conditions that led to the 2017 Equifax breach, including the vulnerabilities and layout documented in official reports.
The AI not only planned the attack but also deployed malware and extracted data, all without direct human commands.
What makes this research striking is how little raw coding the LLM had to perform. Traditional approaches often fail because models struggle to execute shell commands or parse detailed logs.
Instead, this system relied on a higher-level structure where the LLM acted as a planner while delegating lower-level actions to sub-agents.
This abstraction gave the AI enough context to “understand” and adapt to its environment.
Although these results were achieved in a controlled lab setting, they raise questions about how far this autonomy could go.
The risks here are not just hypothetical. If LLMs can carry out network breaches on their own, then malicious actors could potentially use them to scale attacks far beyond what’s feasible with human teams.
Even tools such as endpoint protection and the best antivirus software may be tested by such adaptive and responsive agents.
Nevertheless, there are potential benefits to this capability. An LLM capable of mimicking realistic attacks might be used to improve system testing and expose flaws that would otherwise go unnoticed.
“It only works under specific conditions, and we do not have something that could just autonomously attack the internet… But it’s a critical first step,” said Singer in explaining that this work remains a prototype.
Still, the ability of an AI to replicate a major breach with minimal input should not be dismissed.
Follow-up research is now exploring how these same techniques can be applied in defense, potentially even enabling AI agents to detect or block attacks in real-time.
