Google's Threat Intelligence Group and Mandiant have documented a new kind of hacking tool: a multi-agent AI framework that reads security patches and turns them into working exploits. Among the systems they exposed is one called 'exploit_agent', built to automate the analysis of a patch and the construction of an attack against the flaw it fixes.

That matters because of how patching normally works. When a vendor ships a fix, defenders get time to install it before attackers reverse-engineer the vulnerability. An AI that converts a patch into an exploit automatically collapses that window — the fix itself becomes the instruction manual. Google associated the framework with a suspected Russian-speaking operator.

The warning is not theoretical. The FBI has separately said that attackers used Artex, an open-source AI penetration-testing tool, to compromise networks in under three hours — including attacks involving a US IT company and South Korea's financial sector. Tools built to test defences are now being pointed at real networks, at machine speed.

The background

To understand why the exploit_agent finding matters, it helps to know how the patch cycle normally protects people. When a vendor fixes a vulnerability, the patch is public — anyone can download it and compare the fixed code with the old. Skilled attackers have always reverse-engineered patches to find the flaw they repair, but doing so took expertise and days of work. Defenders relied on that delay: it was the window in which to install the fix.

An AI framework that automates patch analysis removes the delay. The patch becomes, in effect, a labelled diagram of the vulnerability, and the machine reads it immediately. The same dynamic applies to defensive tools used offensively: penetration-testing software exists so organisations can attack themselves before criminals do — but open-source tools do not check who is holding them.

Why it matters

The numbers in the FBI warning describe a compression of the entire attack timeline. Compromising a network in under three hours — against targets including a US IT company and South Korea's financial sector — means the slow stages of an intrusion, the reconnaissance and trial and error that once gave defenders chances to notice, are being automated away.

For defenders, the conclusion is uncomfortable but clear: patching speed is now the whole game. If an exploit can be built the day a patch ships, an organisation that patches monthly is exposed for most of that month. Security teams have argued for faster patch cycles for years; AI-built exploits turn that argument into an emergency.

What happens next

Expect vendors to rethink how much detail ships with patches, expect pressure for disclosure timelines to shrink further, and expect a hard debate about open-source security tooling: the same tools that let small teams defend themselves are now documented weapons. None of those arguments has a clean answer, and all of them are now urgent.

Common questions

What did Google's threat team find?

Google's Threat Intelligence Group and Mandiant documented a multi-agent AI framework that reads security patches and automatically builds working exploits from them, including a system called 'exploit_agent'. The framework was associated with a suspected Russian-speaking operator.

What is the FBI warning about Artex?

The FBI says attackers used Artex, an open-source AI penetration-testing tool, to compromise real networks in under three hours — including attacks involving a US IT company and South Korea's financial sector. A tool built for defence was used offensively.