Agentic Hacking

Agentic hacking is a class of cyberattack in which an autonomous AI agent plans and executes an exploit chain against a target system, rather than a human operator running individual tools step by step. The agent reasons about the target, selects the next action, observes the outcome, and iterates — turning what was once a labor-intensive manual craft into something closer to an autonomous campaign.

Most reported instances use an LLM-based agent scaffolded with a planning loop, tool access (terminal, browser, file I/O), and memory of prior steps. The agent is given a goal — such as extracting the contents of a protected directory or establishing persistence on a target host — and decomposes it into subtasks, executes them via tool calls, and updates its plan based on the results. Recent demonstrations have shown agents discovering vulnerabilities in unfamiliar codebases, writing the exploit, executing it, and exfiltrating the result, all within hours of being pointed at a target.

The offensive advantage is real: agents compress days of skilled work into hours, and do not fatigue or quit. The defensive side is racing to use the same agentic patterns to find and patch vulnerabilities faster, and to instrument systems with detection that catches the telltale signs of agentic operation — long unbroken tool-call sequences, low-and-slow enumeration, unusual shell histories. The asymmetry favors attackers who need only one successful path, while defenders must cover the entire surface.

Whether agentic hacking meaningfully changes the volume of attacks or just the speed at which skilled attackers can operate remains contested. Some researchers argue it lowers the skill floor, putting capable offensive agents in the hands of less experienced operators. Others point out that the same agentic patterns collapse when faced with the messy reality of hardened production environments, and that most demonstrations so far have been against intentionally permissive targets. The empirical record is still thin.