In the last two weeks, the two sides of the offenseβdefense game quietly swapped tables. On November 13, the U.S. FBI issued advisory FLASH-20261113-001, warning that hackers are weaponizing ARTEX, an open-source autonomous penetration-testing framework: it turns natural-language objectives into commands that a swarm of agents executes β reconnaissance, exploitation, privilege escalation, data exfiltration β with some intrusions going from initial targeting to privileged access in under three hours, hitting victims across finance, government, defense, tech, healthcare, energy and telecom. Before that, Norwegian security firm NorForge caught another case: a hacker used a prompt of under 100 characters to direct AI agents across more than 1,200 websites, successfully breaching over 100 of them β an estimated cost of just $50 per successful hack.
What unites these two cases is not another breach, but the fact that the hands doing the work are no longer human. An attack used to require a person finding a weakness, writing a script, running it, failing, tweaking, and re-running β trial and error measured in weeks. Now the agents explore on their own: if the front door is locked they comb the related databases and find a publicly exposed admin panel; when one payload is blocked they try the next, completing hours of trial and error by themselves. The FBI notes attackers also pivot through vendor credentials and trusted business-to-business links β the same playbook as last week's French bank case, where the entry point was an external vendor platform.
The part that should alarm you most is the economics of attacking being rewritten. NorForge's hacker had a total budget of $5,000 and banked over a hundred successful intrusions; researchers counted more than 1,600 publicly exposed attack consoles in early November. Meanwhile model developers are spending tens of billions on compute β attackers only need to pay API usage fees to rent that compute. The bar for attacking is being compressed from a team of specialists to a single person who can write a good prompt; analysts are already calling it the "one-person APT."
For defenders, three things have changed. First, the attackers don't sleep, the defenders do β the FBI says these attacks sharply compress the window defenders have to investigate and contain, and a rhythm of "wake up, then respond" cannot keep up with agents running all night. Second, WAF rules only buy time: SQL injection, auth bypass, file read, template injection β to an agent these are just one line on a trial list; blocked, it moves to the next. Third, the signals themselves have changed β the FBI recommends investigating unexpected LLM API traffic, which means AI-driven attack traffic has become its own new class of indicators to detect.
The correct response on the defense side is to put the same "autonomous" on your own side: inventory every internet-facing system, tighten admin endpoints and application permissions to least privilege (the FBI's concrete advice), and replace "humans watching screens" with 24/7 automated monitoring and anomaly detection β cut traffic and isolate first, then let humans review. If the attack side is already on autopilot 24/7, and defense is still waiting for someone to wake up, it will always be one step behind.
The attack side is already on autopilot 24/7 β the defense side can no longer be the one that sleeps. How LAFA does it: 24/7 AI ops monitoring + WAF / high-defense CDN with automatic defense and traffic switching, isolate first, report second, don't wait for a human to wake up β bringing response speed to the same scale as the agents.