Google sees attackers using AI agents to find and exploit security flaws on their own
Google found an attack system of AI agents that compared software updates, located the fixed flaw and built an attack for it without human help. For defenders, the main lesson is to install updates faster.
In recent months, Google's security team, the Google Threat Intelligence Group, saw attackers leave the whole course of an attack to AI for the first time: from finding a weak spot in software to exploiting it, without a human being involved. Google described this in a report it published .
Google named the system exploit_agent. It consists of several AI agents working together: programs that use a language model to carry out tasks on their own. It can use different language models, both large models over the internet and models the attackers ran themselves. Google does not say exactly which models.
How it worked
When a software maker fixes a security flaw, an update appears. Anyone who puts the old and the new version side by side can see what has changed, and so where the flaw was. The exploit_agent system did this work automatically for popular web programs, such as a widely used extension for WordPress. It then wrote an attack on that flaw and carried it out straight away, on websites that had not yet installed the update.
Google found the system on a poorly secured test server of the attackers. It also held data already stolen from victims, and scripts for guessing passwords and breaking in further. According to Google, the attackers mainly targeted French websites in media, entertainment and education.
Tools anyone can download
Attackers also use freely available programs that are actually meant for testing one's own security, such as ARTEX, CAIRN, CyberStrikeAI and HexStrike AI. These let AI agents search for weak spots and try them out on their own. They are fast, but also noisy: they leave many traces, so a well-monitored network can spot them.
The US federal police, the FBI, warned about ARTEX . The program has been misused since at least August 2026 against companies and governments in several countries, including in finance, technology and healthcare. Some attackers got from choosing a target to administrator rights on a system in less than three hours. The maker removed the program from GitHub, a site where programmers share their code, in October, but copies are still around. In early November, according to the security company Censys, more than 1,600 ARTEX control panels were openly reachable on the internet.
What organisations can do
The main lesson is that the time between an update and an attack is getting shorter. The FBI advises keeping track of which of an organisation's own programs, and those of its suppliers, can be reached from the internet, and opening management screens only to trusted networks. It also helps to monitor login attempts and cloud use, to give accounts no more rights than needed, and to agree in advance how to quickly shut down a suspicious system and revoke passwords. Google itself also uses AI to find flaws earlier and fix them automatically.
What does this mean for the road to the singularity?
AI agents finding security flaws on their own is not new: security companies already use them to test their own software, such as OpenAI with its test agents for websites. What is new is that attackers now also run the whole process automatically. That turns the race between attackers and defenders into a race between machines. Anyone who installs an update only after a week will soon no longer have that week, because an attack on the flaw will be ready within hours.
If that trend continues, updates will have to become just as automatic as the attacks: software that checks, repairs and updates itself without waiting for an administrator. That is good news for small organisations without their own security team, because they would then get the same protection as large companies. It does require those automatic repairs to be reliable, because a faulty update can also cause damage. If that works, software could in a few years be safer than it is now, because flaws disappear faster than attackers can use them.
What is the technological singularity? Read the explanation →